DEV Community: Md Kawsar

Userpilot and the in-app automation field: reaching users the moment they act, no code needed

Md Kawsar — Wed, 21 Oct 2026 13:00:00 +0000

Emails show up hours late, while in-app messages show up as the user is still clicking. Here's how Userpilot stacks up against Appcues, Chameleon, Pendo, and the budget options in 2026.

A user signs up for your SaaS at 2:14 pm. They click around for four minutes, open the settings page, stare at it, and leave.

At 6:00 pm your onboarding email arrives: "Welcome! Here are 5 things you can do." They never open it. The moment is long gone.

That delay between what a user does and how you respond is where most activation dies, and it's exactly why in-app automation tools exist. They talk to users inside the product, while they're mid-task.

I run a small SaaS myself, a size chart maker, so I know this gap well. Here I'll go through Userpilot, which pushed hard into AI and automation this year, and measure it against the rest of the field.

Why timing matters so much

Behavior-triggered messages beat scheduled ones because of context. A tooltip that pops up as someone hovers the export button answers the question they have right now. An email about exporting, sent tomorrow, answers a question they've already dropped.

Most in-app tools can trigger an experience on:

A page or URL the user lands on
A feature they use, or haven't tried yet
A custom event, such as "created first project"
Segment membership, such as "trial users on day 3"
Any mix of the above
Onboarding emails guess what users need. In-app triggers know, because the user just told you with a click.

What Userpilot added in 2026

Userpilot began as an onboarding builder. In 2026 it presents itself as a product growth platform, and three additions stand out, per its 2026 product overview.

Workflows across channels

Userpilot Workflows build behavior-triggered sequences across in-app messages, email, and mobile, and they branch on what users really do. If a user finishes step two in the app, they skip the reminder email. If they don't, the email goes out. Most teams stitch this logic together with three tools and a Zapier account.

Lia, the AI agent

Userpilot's AI agent, Lia, covers five goals: activation, adoption, retention, churn, and monetization. The pitch is that Lia builds reports, breaks results down by segment, runs predictive models, and suggests next steps. It also combines usage data, NPS, surveys, and session replay to flag churn risk months before renewal.

I'm cautiously interested. Predictive churn is only as good as your event tracking, and most teams track less than they think.

Agent analytics

This one is smart. If your product has its own AI agent, Userpilot now tracks those interactions on their own: conversation logs, task completion, and satisfaction. As more SaaS products ship AI features, knowing whether users succeed with them fills a real gap.

The no-code builder in practice

Userpilot's builder runs as a Chrome extension. You open your own app, click an element, and attach a tooltip, modal, slideout, hotspot, or "driven action" that makes the user click something to continue.

Every UI pattern comes with every plan, Starter included, which isn't true of every competitor.

The honest downside: reviewers on Capterra mention occasional interface bugs, a limited template library, and a learning curve on the analytics side. It's no-code, though it still takes thought.

The in-app automation field, side by side

Here's how the main options compare in 2026. Prices change often and vary by source, so use these as starting points to verify rather than quotes.

Userpilot

Starts around $249 to $299 a month for about 2,000 monthly active users, depending on source and billing. Best for product-led SaaS teams that want onboarding and analytics together. Its analytics (autocapture, funnels, cohorts, session replay) run deeper than most direct competitors.

Appcues

Starts around $249 a month for 2,500 MAUs. It's the quickest to set up, often live within hours, and a great fit for non-technical teams. Analytics are lighter, and pricing can jump at higher tiers.

Chameleon

Starts around $279 a month. Strong at making experiences look native to your product, with CSS control and embedded patterns. Setup takes longer, and it relies on external analytics.

Pendo

Enterprise pricing, roughly $15,000 to $140,000+ a year by quote, according to Chameleon's comparison. Very strong analytics and roadmapping, and overkill for most startups.

Budget picks

UserGuiding sits around $89 to $389 a month and handles basic onboarding well. Userflow runs roughly $240 to $800 a month with a clean builder. Both make solid starting points if you have under a few thousand users.

The cheapest tool that gets a user to their first win is the right tool. Features you never use add no value.

A simple way to choose

I don't think there's one winner. This is how I'd pick, based on what your team really needs.

Pre-product-market-fit with a small user base: start with UserGuiding or Userflow. Learn which triggers matter before you pay for analytics.
Need something live this week with no developer time: Appcues.
Brand and UI polish matter a lot: Chameleon.
Want onboarding, analytics, surveys, and multi-channel workflows in one tool: Userpilot.
Enterprise, with a big analytics budget and a product ops team: Pendo.

Five in-app automations worth building first

Whichever tool you pick, these are the flows I'd build before anything fancy. Each one triggers on behavior, never on a timer.

The stuck-on-setup nudge. Trigger: a user visits the setup page twice without finishing. Show a short checklist with the one next step.
The first-win celebration. Trigger: a user completes the core action for the first time. Show a small success modal and point to the next valuable feature.
The unused-feature tooltip. Trigger: a user has been active for 7 days but hasn't touched a key feature. Show one tooltip on that feature, one time.
The pre-churn check-in. Trigger: usage drops by half over two weeks. Show a one-question survey: "What's getting in the way?"
The upgrade moment. Trigger: a user hits a plan limit. Show the upgrade offer right there, with what it unlocks.

That fifth one is where in-app beats email by a mile. Someone who just hit a limit is at peak intent. An email tomorrow interrupts them, while a message right now helps.

Mistakes I'd avoid

Too many popups. If users close three modals in their first minute, you've taught them to close everything.
Tours nobody asked for. Ten-step product tours on first login have terrible completion rates. Short and contextual wins.
No frequency caps. Limit how many experiences a user can see in one session.
Tracking nothing. If you can't measure whether a flow changed activation, the tool is wasted money.
Forgetting small screens. A tooltip that looks great on a 27-inch monitor can cover the entire button on a laptop. Preview every flow at small sizes before you publish.
Never retiring old flows. Products change. A tour pointing at a button you moved three months ago is worse than no tour.

How to tell if it's actually working

This is the part most teams skip, and it's why so many in-app tools get cancelled after a year: nobody can prove they did anything.

Before you launch a single flow, pick the numbers you'll watch. I'd start with these four:

Activation rate. The share of new signups who reach your core "aha" action within a set window, such as 7 days. In-app automation should move this one most.
Time to first value. How long it takes to get from signup to that first win. Good onboarding shortens it.
Flow completion rate. For each checklist or tour, how many people finish. Low completion means it's too long or appears at the wrong moment.
Adoption of the feature you're promoting. If you nudge users toward exports, did export usage actually rise?
Then run flows as experiments. Show the new experience to half of new users and hold it back from the other half for two to four weeks. Most tools in this category support some form of A/B testing or control groups. If activation doesn't move, the flow isn't helping, however nice it looks.

One more tip from my own habit with data: write down what you expect before you look at the results. It keeps you honest. After the fact, it's very easy to call a 1% bump "the plan all along."

A flow that can't prove it moved activation is decoration, nothing more.
The bottom line
In-app automation comes down to timing. The whole category exists to close the gap between what a user just did and when you respond.

Userpilot is the most complete mid-market option right now, especially with multi-channel Workflows and agent analytics. Still, Appcues is faster, Chameleon is prettier, and the budget tools are plenty for early teams.

Pick based on the first flow you need to build. Then build it, measure it, and upgrade only when the tool, not your plan, is what's holding you back.

Enjoyed this read? Follow me for more stories like it. You can also reach me on LinkedIn or Telegram @kawsarlog for updates on my work.

I take on freelance scraping projects, so if you need data collected, message me on LinkedIn or email hi@kawsarlog.com

AI model release trackers: how to keep up with 400+ models without losing your mind

Md Kawsar — Tue, 20 Oct 2026 15:00:00 +0000

171 models came out in 2026 before October, and 26 of them arrived in September alone. Nobody can test them all, so here are the trackers worth saving and an easy system for choosing which releases deserve your time.

Just look at the final ten days of September 2026. xAI released Grok 4.7. Anthropic released Claude Opus 5.5, followed by Sonnet 5.5 a few days later. OpenAI rolled out new GPT-6 Sol variants. Every one of them came with charts claiming it was the best so far.

I'd barely finished one launch post before the next one started trending.

If it feels impossible to keep up, that's because it nearly is. LLM Gateway's release timeline now counts 416 models from 54 providers since 2022. 171 of them came out in 2026, and 26 arrived in September alone, from 16 separate providers.

No one can test all of that, including you, me, and the people sharing benchmark charts on launch day. So this post does two jobs: it shows you the AI model release trackers worth saving, and it hands you an easy filter for spotting which releases matter to your work.

Why keeping up got so hard

Three years ago, a big model launch was an event. Today it's more like the weather.

Here's why the pace took off:

More labs. On top of the big US names, there's Qwen, DeepSeek, Moonshot's Kimi, Z.ai's GLM, Xiaomi's MiMo, Mistral, and plenty more, each releasing on its own timeline.
More sizes. A single "release" is often a whole family: a flagship, a mid-range model, a quick cheap model, and sometimes a coding version.
Point releases. Labs now release 5.5, 5.6, and 5.7 rather than waiting for a big version jump.
Price changes count too. When a model gets 80% cheaper, as OpenAI's GPT-5.6 Luna did on July 30, that can matter more than a brand-new model.
Forget knowing every model. Know the three that matter for your work, and notice when one of them gets replaced.

The trackers worth bookmarking

These are the free trackers I'd use, depending on what you want to follow.

LLM Gateway timeline: the biggest list

LLM Gateway covers 416 models from 54 providers and 33 model families, with release counts for every month. It's the quickest way to see "what came out this month?"

Use it for: a big-picture view of the entire market.

Evertune: the clean summary

Evertune's AI Model Release Tracker is more focused, with about 133 releases and updates from six major providers (OpenAI, Anthropic, Google, Meta, DeepSeek, and others). Each entry has a date, a plain-English summary, and a link to the official announcement, and it's updated every day.

Use it for: getting a quick, hype-free read on what a major release changed.

Opper, BenchLM and LMMarketCap: quick checks

Opper's release tracker, BenchLM's live updates, and LMMarketCap's tracker are handy for checking the past 24 to 48 hours, with benchmark or pricing context shown beside each release.

Use it for: "did something big drop today?"

Pricing trackers: the underrated one

Releases grab the headlines, but changes to price and limits are what hit your budget. UsagePricing's activity log records dated pricing changes across AI and SaaS vendors.

Use it for: spotting cost changes before your invoice shows them.

The three-question filter

This is the filter I use to decide whether a new release gets any of my time. It takes around 30 seconds.

Is it from a provider I already use, or a direct replacement for a model I use? If not, move on unless the next two answers are a strong yes.
Does it change price, speed, or limits by at least 20%? **Tiny benchmark gains rarely matter for real work. Big shifts in cost or speed do.
**Does it unlock something I couldn't do before? That could be longer context, a new modality, stronger tool use, or a much higher coding score on tasks like mine.

If all three answers are no, I close the tab. That one rule cuts out most of the noise. For most people, only a few releases each month will get through.

Benchmarks show how a model did on someone else's test. Your own five-task test shows how it'll do for you.

Build your own five-task test

When a release gets through the filter, skip the launch-day charts and run your own quick test. Here's how to build one once and reuse it forever.

Pick five real tasks from your work. Mine include turning a messy HTML page into JSON, writing a Python function for pagination, summarizing a client brief, cleaning a CSV with mixed date formats, and drafting a short reply to a client.
Save the exact inputs. Use the same inputs every time so the comparison stays fair.
Save your current best outputs **to use as your baseline.
**Score simply: better, same, or worse, along with cost and time.
Keep a log. Give each model a row and each task a column.

After a few rounds, your log will be more useful than any public leaderboard, since it's based on your work.

A weekly routine that takes 15 minutes

Here's the routine I'd recommend for anyone who builds with AI but doesn't want keeping up to turn into a second job.

Monday, 5 minutes: scan one big tracker for last week's releases and run each through the three-question filter.
Monday, 5 minutes: look at a pricing tracker for changes to the models you pay for.
Once a month, 30 minutes: run your five-task test on any release that got through the filter.
Once a quarter: go over which models you pay for and whether a cheaper one can now do the same job.
That's all. You won't know everything, but you'll know what matters.

Keep a "my models" sheet

Public trackers cover the entire market. You also need a small tracker for your market, meaning the models you actually use. Mine is a basic sheet with one row for each model and these columns:

Model name and provider
What I use it for (scraping parsers, writing, coding, summaries)
Current price per million input and output tokens
Where I use it (API, a subscription, a tool like Cursor)
Known deprecation or retirement date
The date I last tested it
The cheaper model I'd switch to if prices rose
Setting it up takes 20 minutes. From then on, each release faces one easy question: does it beat anything in my sheet? And when a provider announces a retirement, like Moonshot ending its Kimi K2.5 and moonshot-v1 API family on August 31, you'll know right away whether it affects you.

Open models deserve a separate glance

Many of those 171 releases in 2026 are open-weight models from labs such as Qwen, DeepSeek, Z.ai, Moonshot, and Mistral. If you only use chat apps, it's easy to overlook them. I'd still check them briefly each month, for two reasons.

First, price. Open models hosted by inference providers are often far cheaper for bulk jobs like classification, extraction, and cleanup. In scraping pipelines that handle thousands of pages, the savings add up quickly.

Second, control. You can run many of them on your own hardware, fine-tune them, and keep data inside your own environment. For some clients, that's the deciding factor.

There's no need to track every open release. Just keep one open model in your five-task test so you always know what the budget option can handle.

Signals that a release matters

Over time, I've spotted a few signs that separate real changes from noise:

Tool builders switch fast. When coding tools and API gateways add a model within 48 hours and set it as a default, take notice.
The price drops at the same quality. Opus 5.5 launching at $4 input and $20 output per million tokens, down from Opus 5's $5 and $25, is the sort of change that moves real budgets.
Old models get deprecated. Once your current model gets a retirement date, you have to pay attention, like it or not.
Developers you trust actually change their stack, beyond just tweeting about it.

What to ignore

Launch-day leaderboard screenshots that skip the methodology
"This changes everything" posts within the first six hours
Benchmarks on tasks that look nothing like what you do
Small version updates from providers you don't use
Call it self-defense. It's the only way to stay sane when a new model comes out almost every day.

You don't need to see everything

An AI model release tracker can't help you keep up with 400+ models, and nothing else can either. What it can do is give you a fast, reliable look at what changed, so you can run a simple filter and move on.

Save one big tracker and one pricing tracker. Build your five-task test. Give it 15 minutes a week.

You'll miss lots of releases. That's fine. You'll catch the ones that matter for your work, and that's the only kind of keeping up that pays off.

Enjoyed this read? Be sure to follow me *for more stories that make people feel things. You can also connect with me on *LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

Physical AI for small teams: what warehouse, lab and farm robots can do for you

Md Kawsar — Mon, 19 Oct 2026 13:01:00 +0000

Robots used to be something only big companies could budget for. Cheaper edge chips, rental plans, and AI that picks up tasks faster are bringing them within reach of teams of 5 to 50. Here's what's real, what it costs, and how to begin.

For most of my life, the word "robot" made me think of a car factory: huge orange arms behind safety cages, owned by companies with billion-dollar budgets.

That image is going out of date quickly. In 2026, a small farm can rent a robot that trails pickers down the rows hauling the harvest. A small biotech lab can buy a liquid-handling robot for roughly the cost of a used car. And a 3PL warehouse with 20 employees can lease picking robots month to month.

The industry's name for this is physical AI: AI that steps out of the chat window to see, move, and do real work in the physical world. MarketsandMarkets estimates the market at around $1.5 billion in 2026, rising to more than $15 billion by 2032.

Those are big numbers. I'm more interested in a smaller question: what does physical AI mean for a small team?

Why this is happening now

Three shifts happened all at once.

The brains got cheap. In 2024, NVIDIA's $249 Jetson Orin Nano Super made serious edge AI affordable. Its follow-up, the Jetson Orin Nano 2, announced in August 2026, promises twice the inference at the same entry level. Robots can now see and make sense of a scene right on the device.

Robots got easier to teach. Vision-language models help robots follow instructions and cope with messy spaces, so every movement no longer has to be programmed by hand.

You don't have to buy them. Robots-as-a-service (RaaS) lets you rent robots by the month or pay per task, much like software shifted to subscriptions.

Physical AI arrived as three small steps landing together: cheaper brains, easier training, and renting in place of buying.

What robots actually cost in 2026

First, the numbers. These ranges come from Standard Bots' 2026 price guide and vendor pages.

Collaborative robot arms (cobots): $25,000 to $75,000 base, $40,000 to $150,000 as a full system
Mobile robots and AGVs: $25,000 to $150,000 base, plus $5,000 to $15,000 per unit for fleet software
AI service robots: $30,000 to $200,000+
Lab liquid handlers: the Opentrons Flex starts around $26,400
Farm platform robots: Burro's autonomous carts run roughly $25,000 to $85,000
Humanoids: $150,000 to $1,000,000+, still mostly for research and pilots
According to Standard Bots, most robots pay for themselves in 18 to 30 months. That's an average across lots of use cases, so your own numbers matter more. Keep in mind that AI features add 20% to 40% to the base price, and integration can tack on another $10,000 to $50,000.

Warehouses: the most mature use case

If you run a small warehouse, a 3PL, or an ecommerce fulfillment operation, this is where physical AI has the strongest track record.

Autonomous mobile robots (AMRs) from companies like Locus Robotics work side by side with pickers. Rather than a person pushing a cart for miles each day, robots bring totes to workers and take picked items away. The robot does the traveling, and the person does the picking.

What it means for a small team:

With RaaS contracts, you can add robots for peak season and scale back afterward
The biggest gain is less walking, often the largest hidden cost in picking
New hires get up to speed faster, because robots lead them to the right shelf
What to watch: your warehouse management system needs to connect with the robots. If your inventory lives in spreadsheets, sort that out first.

Labs: automation for the small bench

Lab automation once meant giant machines inside pharma companies. Today, a small university or startup lab can set a liquid-handling robot right on a bench.

Take the Opentrons Flex. It pipettes from 1 to 1,000 microliters, has a gripper that moves plates, and runs protocols you build with a drag-and-drop designer, a Python API, or an AI tool called OpentronsAI. Labs in more than 40 countries use it for tasks like PCR setup, DNA extraction, and sequencing prep.

What it means for a small team:

Repetitive pipetting shifts from people to machines, cutting both errors and strain
Experiments can run overnight with nobody watching
Protocols turn into shareable files, which makes results easier to repeat
What to watch: a robot can only be as good as its protocol. Set aside time for someone to build and test them.

Farms: robots that work outdoors

Farming is where physical AI gets tough, because fields are unpredictable: bumpy ground, shifting light, mud, and plants that grow differently each season.

That's why today's most successful farm robots stick to narrow, clearly defined jobs. Burro builds small self-driving carts that use cameras, AI, and precise GPS to follow crews, haul harvested produce, scout crops, or pull loads. At the larger end, Carbon Robotics' LaserWeeder uses computer vision to spot weeds and burn them with lasers, without any herbicide.

What it means for a small team:

Harvest labor shortages are the main reason, and carts that carry crates let pickers keep picking
Scouting robots turn a walk through the field into data, showing which rows are stressed and which need care
Leasing and co-ops can split the cost among several small farms
What to watch: how reliable it is in your conditions. Ask for a demo on your land instead of a showcase field.

For small teams, the robots that pay off are the ones doing one dull job reliably every single day, however smart the flashier ones look.

What physical AI does to your data

I think this part gets missed, and it's the part closest to what I do.

Each of these robots also works as a data collector. A picking robot knows how long every order took and where people waited. A lab robot records every step of every run. A farm robot maps each row and logs what it saw.

For a small business, that's a brand-new asset. All of a sudden you have exact, timestamped data on your operations that used to exist only in someone's memory. Used wisely, it shows where your process drags, which products are tricky to pick, or which parts of a field are falling behind.

Ask two questions before you sign any RaaS contract:

Who owns the data the robot collects?
Can I export it in a usable format, like CSV or through an API?
If you hear "the vendor" and "no," push back. That data could be worth as much as the work the robot does.

The simple math before you call a vendor

You can figure out whether a robot might make sense without hiring a consultant. A quick back-of-the-envelope calculation shows whether it's worth a call.

Picture a small warehouse where three pickers each spend around half their shift walking. At $20 an hour with overhead, eight-hour shifts, and roughly 250 working days, all that walking costs about** $60,000 a year**. If robots cut the walking in half, you'd free up about $30,000 a year in labor time for packing, receiving, or customer work.

Next, compare that with a RaaS contract for the number of robots you'd need. If the yearly rental comes in well below the value of the time you'd free up, run a pilot. If the two are close, softer benefits like fewer injuries and easier hiring will decide it. If the rental is much higher, give it a year or two, since prices keep dropping.

These numbers are just examples to show the method, so swap in your own. The point is that you can work this out on a napkin before any vendor sends a sales deck.

A practical way for a small team to start

If you're curious but don't want to risk the business, this is how I'd go about it.

List your most repetitive physical task. Think walking, carrying, pipetting, or sorting. The duller the task, the better it suits a robot.
Measure it for two weeks. Track hours spent, mistakes made, and any injuries or strain. You need a starting point.
Talk to two or three RaaS vendors. Ask about minimum contract length, uptime guarantees, repair response times, and data ownership, as Layer3 Labs' RaaS guide recommends.
Pilot one robot for one task. Keep the contract short and the success measure clear.
Plan for people, not replacement. The best outcomes come when people shift to work the robot can't handle, rather than being let go.

My honest prediction
I don't expect small teams to fill up with humanoid robots any time soon. Humanoids are still pricey, early-stage, and mostly stuck in pilot projects.

What I do expect in the next two or three years is quieter: carts trailing farm crews, robots hauling totes around small warehouses, bench robots pipetting through the night, and cameras with cheap AI chips watching a production line for problems.

None of that will grab headlines. Yet for a 20-person business, a robot that spares one employee five miles of walking a day is a big deal. That's what physical AI for small teams really looks like: fewer boring hours, with no science fiction required.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

7 marketing automation tools made for the way SaaS works

Md Kawsar — Sun, 18 Oct 2026 15:00:00 +0000

Mailchimp is built around lists. SaaS runs on events, trials, and accounts. These seven tools were designed for SaaS, and here's which one suits your stage.

Sooner or later, every SaaS founder tries running lifecycle email through a newsletter tool. It works well for about three weeks.

Then you want to email trial users who still haven't connected their data. Or message every account where the admin invited teammates and nobody logged in. Or halt the upgrade sequence as soon as someone pays in Stripe. Your newsletter tool can't help, because it understands lists and has no idea about behavior.

SaaS works differently. Your users take actions inside a product, and your marketing should respond to those actions. That calls for tools built around events as well as contact fields.

Below are seven marketing automation tools for SaaS made for exactly that job, with honest notes on cost and fit.

What "built for SaaS" actually means

Before the list, this is the test I apply. A tool is truly made for SaaS if it does most of these well:

Event-based triggers: "user created first project" can start a workflow
Product data sync: native connections to Segment, your database, or an API
Billing awareness: Stripe events like trial ending, payment failed, upgraded
Account-level data: for B2B, knowing that five users belong to one company
Transactional plus marketing email in one place
Newsletter tools tend to cover one or two of these. The tools below cover most or all of them.

Newsletter tools ask "who's on this list?" SaaS tools ask "what did this user just do?" That single difference changes the whole game.

1. Customer.io: the powerful default

Customer.io is where many growth teams land once lifecycle messaging gets serious. It's designed around events and profiles, with a visual workflow builder and channels beyond email, such as push, in-app, SMS, and WhatsApp.

Pricing: Essentials begins at $100 a month for 5,000 profiles and 1 million emails. Premium begins at $1,000 a month, billed yearly. A startup program gives companies that have raised under $10M 12 months free, so check that before paying anything.

Watch out for: all that power needs setup time. You'll want someone who's comfortable with event data.

Best for: product-led SaaS with solid event tracking and plans for several channels.

2. Loops: the modern, simple one

Loops is a popular choice for early-stage SaaS teams in 2026, because it feels like it was made by people who can't stand bloated email tools. It has a clean editor, event triggers, and transactional email included on paid plans.

Pricing: free for up to 1,000 subscribed contacts and 4,000 sends in any rolling 30 days, with every feature included. Paid plans are priced by subscribed contacts rather than sends, starting near $49 a month. Team seats are free, and unsubscribed contacts aren't counted.

Watch out for: it has fewer advanced features than Customer.io and is a weaker fit for complex B2B account logic.

Best for: early SaaS teams and indie founders who want product and marketing email in one tidy tool.

3. Encharge: B2B SaaS on a budget

Encharge hits a sweet spot, with a visual flow builder, built-in Segment and Stripe integrations, event-based segmentation, and company profiles for account-based messaging.

Pricing: its pricing page shows plans from about $99 a month, though some comparisons mention cheaper starting points. Check the current tiers for your contact count.

Watch out for: its ecosystem is smaller than the big names'. That's fine for most teams, but confirm your key integrations are there.

Best for: B2B SaaS teams that want Stripe and Segment hooked up without paying enterprise prices.

4. Userlist: account-level automation done right

Userlist concentrates on something competitors often get wrong: B2B SaaS where users belong to companies. It handles company-level automation and even many-to-many relationships, where a single user can be part of several accounts.

Pricing: plans cover 10,000 users, and extra users cost $10 per 1,000 on Basic or $12 on Professional. Early-stage startups with less than $500K in funding can get 50% off. Every plan comes with marketing, lifecycle, and transactional email, plus in-app messages.

Watch out for: it offers fewer channels than Customer.io and focuses mainly on email.

Best for: B2B SaaS where the buyer is the whole account rather than one person.

5. Userpilot: when the message belongs inside the product

Some lifecycle messages work better outside the inbox. Userpilot covers in-app onboarding, tooltips, checklists, and surveys. In 2026 it added multi-channel Workflows that mix in-app, email, and mobile, along with an AI agent called Lia that spots churn risk.

Pricing: begins at roughly $249 to $299 a month for around 2,000 monthly active users, depending on how you pay.

Watch out for: it works alongside your email tool and won't replace it.

Best for: teams whose biggest challenge is getting users activated inside the product.

6. Intercom: conversations first

Intercom brings together live chat, support, product tours, and outbound messages. Its AI agent, Fin, answers support questions for $0.99 per outcome. A big 2026 update: the company changed its name to Fin in May, and Salesforce completed its acquisition in September. The helpdesk product keeps the Intercom name.

Pricing: seats begin at about $49 a month, plus AI usage.

Watch out for: its marketing automation is lighter than that of dedicated tools, and nobody knows yet what pricing will look like under Salesforce when you renew.

Best for: SaaS teams whose growth comes from support and onboarding conversations.

7. HubSpot Marketing Hub: the one you graduate to

HubSpot wasn't made only for SaaS, but it makes the list because so many SaaS companies move to it once a sales team joins. It ties marketing, CRM, and sales together, with its Breeze AI tools on top.

Pricing: the CRM is free to start. Marketing Hub Professional runs about $800 a month for 1,000 contacts, and custom event-based features are on the higher tiers.

Watch out for: it handles product events less smoothly than Customer.io, and costs rise fast.

Best for: SaaS with a proper sales team, where marketing and sales need one shared system.

The right tool is the one your team will keep feeding with data, however powerful the alternatives are.

Which one fits your stage

Here's an easy way to choose.

Pre-revenue or under $10K MRR: Loops. It's free to start, easy to use, and handles both transactional and marketing email.
Growing product-led SaaS with good event data: Customer.io. Look into the startup program first.
B2B SaaS where accounts matter: Userlist or Encharge.
Activation is the bottleneck: add Userpilot alongside your email tool.
Support conversations drive the business: Intercom.
Sales-led with a real CRM need: HubSpot.

Five SaaS automations to build first

Whichever tool you choose, begin with these. They cover most of the lifecycle, and every one runs on events.

Activation nudge. Trigger: the user signed up but hasn't taken the key action within 48 hours. Send one short email with a single link to that action.
Trial ending. Trigger: 3 days before the trial ends. Focus on what they've built, and keep the price in the background.
Failed payment. Trigger: a Stripe payment failed. Keep it friendly and quick, with a direct link to update payment. This one saves real revenue without much fuss.
Team invite follow-up. Trigger: an invite went out and the invitee hasn't joined after 3 days. Remind the admin as well as the invitee.
Win-back. Trigger: no login for 30 days. Ask one honest question: what got in the way?

Get your event tracking ready in a weekend
Each tool here is only as smart as the events you feed it. Here's an easy way to prepare before you choose one.

Write down 8 to 12 events, no more. For example: signed_up, onboarding_completed, created_first_project, invited_teammate, connected_integration, hit_plan_limit, trial_started, subscription_started, payment_failed, cancelled. That list covers nearly every automation you'll build in your first year.

Add a few properties to each. Plan name, company ID, and a timestamp are usually plenty. Don't try to track everything. Messy tracking plans are a big reason teams abandon lifecycle marketing.

Send them from one place. Either a customer data tool like Segment, or have your backend call the email tool's API. Hook Stripe up directly for billing events.

Test with your own account. Sign up, click around, and watch the events show up in the tool. If your own journey isn't clear to you, your automations won't see it clearly either.

Speaking as a data person, I'll be blunt: clean events beat clever workflows every time.

Mistakes to skip

Buying before tracking. Without clean events, no tool can help you. Get tracking in place first.
Too many emails in week one. Five emails in five days comes across as desperate. Let user behavior set the pace.
Ignoring deliverability. Authenticate your domain and keep your lists clean, or your lifecycle emails end up in spam beside cold outreach.
Separate tools for transactional and marketing that share no data. That's how users end up getting a "welcome" email after they've cancelled.

Events first, tools second

The best marketing automation tools for SaaS all share one idea: respond to what users do, on top of who they are.

Begin with the simplest tool that handles events for your stage. Build the five automations above. Track activation and retention rather than open rates. And only upgrade once the tool itself, and not your data, is what's holding you back.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

The 12 best AI sales automation tools for 2026: from Apollo and Clay to a full outbound stack

Md Kawsar — Sat, 17 Oct 2026 13:00:00 +0000

A layer-by-layer tour of a modern outbound stack: which tools own each layer, what they cost, and which ones a small team can skip.

Most "best sales tools" lists are logos in random order, which isn't how anyone buys.

Real outbound teams buy in layers. First you need data. Then you clean and enrich it. Then you send. Then you handle replies and calls. Every tool here sits in one of those layers, and the mistake I see most is paying for three tools that do the same job.

I build lead data for clients, so I spend most of my time at the bottom of this stack. Here are the 12 AI sales automation tools I'd look at in 2026, arranged the way you'd assemble them.

The outbound stack at a glance

Before the list, here's the map. A modern outbound stack has five layers:

Data: who you contact
Enrichment and research: what you know about them
Sending: getting messages out without landing in spam
Engagement and conversations: sequences, calls, replies
Agents: AI that handles chunks of the above by itself
A three-person startup might cover all five with two tools, and a 50-rep team might run eight. Both can be right. What matters is that each layer gets covered exactly once.

Stack by job, not by hype. Two tools doing the same job means one of them is a subscription you forgot to cancel.

Layer 1: data

1. Apollo.io

Apollo is where most small teams start, and there's good reason for it. It pairs a large contact database with sequences, a dialer on higher plans, and AI-assisted writing, and the free plan is real.

Prices per Warmly's 2026 breakdown: Basic is $59 per user per month ($49 annual), Professional is $99 ($79 annual), and Organization is $149 ($119 annual, with a 3-user minimum). Watch the credit system, since mobile numbers and exports each draw on separate credits.

Best for: teams that want data and sending together without a big budget.

2. ZoomInfo

The enterprise heavyweight. It offers deeper company data, org charts, intent signals, and strong North American coverage. Pricing is custom, usually an annual contract in the five figures.

Best for: mid-market and enterprise teams where accuracy on large accounts justifies the cost.

Layer 2: enrichment and research

3. Clay

Clay is the tool that turned "GTM engineer" into a job title. Picture a spreadsheet where each column can call a data provider, an AI model, or an API. Its AI research agent, Claygent, can browse websites and pull out answers like "does this company mention SOC 2 on its site?"

Clay changed its pricing on March 11, 2026, splitting billing into Data Credits and Actions. According to Databar: Free (100 credits, 500 actions), Launch at $185 a month, and Growth at $495 a month. Marketplace data got 50% to 90% cheaper, but HTTP API calls now cost Actions and need the Growth tier. If you rely heavily on custom API calls, redo your math before you renew.

Best for: teams that want custom, multi-source enrichment and have someone technical to build it.

4. Common Room

Common Room gathers buying signals from your website, product usage, GitHub, community channels, and social activity, then ties them to accounts. Rather than hunting for strangers, it spots who's already warm.

Best for: product-led and developer-focused companies with an active community.

Layer 3: sending

5. Instantly

An all-in-one cold email tool with unlimited sending accounts, unlimited warmup, a built-in lead database called SuperSearch (450M+ records), inbox placement testing, and a light CRM. Plans run $47 to $358 a month.

Best for: in-house teams that want sending and data in one polished tool.

6. Smartlead

The agency favorite. It offers unlimited mailboxes, a strong API, per-client workspaces, and white-label dashboards. Plans run $39 to $379 a month, with placement testing and white-label sold as add-ons.

Best for: agencies running outbound for many clients. (I wrote a full Smartlead vs Instantly comparison if you can't decide between the two.)

7. lemlist

lemlist leans into personalization and multichannel work: email plus LinkedIn steps, personalized images, and a growing lead database. Pricing runs roughly $59 to $159 per seat per month, according to recent 2026 breakdowns.

Best for: small teams that want email and LinkedIn in one sequence and care about creative personalization.

Layer 4: engagement and conversations

8. Outreach
A sales engagement platform for larger teams, covering sequences, forecasting, deal management, and AI that summarizes calls and suggests next steps. Pricing is custom.

Best for: sales orgs where managers need visibility and process across many reps.

9. Salesloft

Outreach's long-time rival, now combined with Clari after their merger, which adds revenue forecasting to the engagement side. Custom pricing.

Best for: teams that want engagement and forecasting from one vendor.

10. Gong

Gong records and analyzes sales calls, then turns them into coaching, deal risk alerts, and forecasts. It's the best-known conversation intelligence tool for good reason. Pricing is custom and usually combines a platform fee with per-seat licenses.

Best for: teams where calls drive deals and managers want real coaching data.

Layer 5: AI agents

11. HubSpot Breeze Prospecting Agent

If you already live in HubSpot, the Prospecting Agent researches accounts, watches signals, and drafts outreach. Since April 2026 it costs $1 per lead it recommends. Reviews agree the drafts need a human check, so treat it as a research assistant with a pen.

Best for: HubSpot teams that want AI research without adding another vendor.

12. AI SDR platforms (11x, Artisan, Amplemarket)

This is the newest and loudest category: "AI SDRs" that promise to find leads, write emails, send them, and handle replies. 11x and Artisan sell named AI workers, while Amplemarket blends a data platform with AI-driven outreach.

I'll be honest: I'm skeptical of fully autonomous outbound. From my experience with lead data, teams that get results use these tools for research and first drafts, then keep a human in charge of tone and targeting. Pricing is usually custom and annual, so ask for a short pilot before you sign.

Best for: teams with clean data and a clear ICP who want to test AI doing more of the top of the funnel.

AI can write a thousand emails an hour. It still can't tell you which thousand people should get one. That decision is yours.

Three stacks for three kinds of teams

Here's how I'd combine these at different budgets.

The solo founder (under $150 a month)

Apollo free or Basic for data
Instantly Growth for sending
A simple email verifier before every send
That's the whole stack. Skip Clay until you've outgrown Apollo's filters.

The 3 to 10 person sales team ($500 to $2,000 a month)

Apollo Professional for data and dialing
Clay Launch for custom enrichment
Instantly or Smartlead for cold email
HubSpot as the CRM, with Breeze Assistant switched on

The agency

Clay for research and list building
Smartlead with white-label for sending
Your own data sources or scraped lists for niches the databases miss

That last bullet is my home turf. The big databases do well for US tech and SaaS, then thin out quickly for local businesses, non-English markets, and niche industries. In those cases, custom scraped data is often the only route to a list that isn't the same one your competitors are emailing.

A quick word on data quality, since it's the layer I know best. Every database on this list holds stale records, because people change jobs, companies rebrand, and domains expire. Industry estimates for B2B data decay often land around a quarter or more of records per year. That's why a list pulled six months ago can bounce far more than a fresh one. Whatever you use, re-verify before every campaign instead of once a year.

How to avoid overbuying

A few rules I'd follow before adding any new sales tool:

Name the layer it covers. If a tool already fills that layer, you need a very good reason.
Run a 30-day pilot around one metric. Meetings booked, reply rate, or hours saved. Pick one.
Check the credit math, not the sticker price. Apollo, Clay, and HubSpot all bill on credits that can surprise you.
Protect your domains. No tool can repair a burned sending reputation.
Verify data before sending. Every database has stale records, and bounces hurt everyone on your domain.

Start small, aim well

The best AI sales automation tools in 2026 are the ones that fit cleanly into one layer of your stack and do that job well. The flashiest AI demos matter far less.

Start small. Apollo plus a sending tool takes most teams surprisingly far. Add Clay when you need custom research, Gong when calls matter, and AI agents only once your data and targeting are already solid.

Tools multiply output, but they can't fix aim. Get the aim right first.

Enjoyed this read? Follow me for more stories like it. You can also reach me on LinkedIn or Telegram @kawsarlog for updates on my work.

I take on freelance scraping projects, so if you need data collected, message me on LinkedIn or email hi@kawsarlog.com

I reviewed HubSpot Breeze AI: Assistant, Customer Agent, Prospecting Agent and Intelligence

Md Kawsar — Tue, 13 Oct 2026 15:00:00 +0000

HubSpot now bills for its AI agents only when they deliver: 50 cents per resolved chat and $1 per qualified lead. Here's what each part of Breeze does, what it costs, and where it comes up short.

Most AI pricing charges you for trying. You buy seats, burn through tokens, and hope something useful comes out the other end.

In April 2026, HubSpot turned that around. Its Customer Agent now costs $0.50 per resolved conversation, and its Prospecting Agent costs $1 for each lead it recommends for outreach. When the agent hands a chat to a human, that chat is free. HubSpot's chief customer officer Jon Dick told MarTech: "You pay when it works, full stop."

That's a big promise, so I went through every part of the HubSpot Breeze AI suite to separate what's real from what's marketing, and to work out who should actually switch it on.

What Breeze covers

Breeze is the label HubSpot puts on all the AI spread across its CRM, so it's several products under one name. Four pieces matter most:

Breeze Assistant (formerly Breeze Copilot): the chat sidebar inside HubSpot
Customer Agent: an AI support agent that handles tickets and chats
Prospecting Agent: an AI that researches accounts and drafts outreach
Breeze Intelligence: data enrichment and buyer intent
HubSpot also has newer agents, such as a Content Agent and data agents, but most teams will spend their time and money on these four.

Breeze only knows what your CRM knows. That's the whole appeal and the main limit.

Breeze Assistant: the free piece everyone should try

HubSpot renamed Breeze Copilot to Breeze Assistant, and it's the easiest win in the suite. Every HubSpot edition includes it, even the free plan, and it uses no credits.

What it does well:

Summarizes a contact, company, or deal record before a call
Drafts follow-up emails using the record's context
Prepares meeting notes
Handles small CRM actions, like creating tasks or updating properties
The catch is simple, and Sybill's review put it well: "garbage in, breezy garbage out." If your reps log two-word notes, you'll get a two-word summary stretched into a paragraph.

My take: switch it on today. It's free, and it will quietly show you how messy your CRM data is.

Customer Agent: the strongest piece

HubSpot is proudest of this agent, and its numbers are solid. The company says the Customer Agent resolves 65% of conversations and shortens resolution time by 39%, across 8,000 customers.

It reads your knowledge base, help docs, and past tickets, then answers customers on chat and other channels. When it isn't sure, it passes the conversation to a person.

Pricing is the interesting part:

Before April 14, 2026: $1.00 per conversation
After: 50 credits, about $0.50, per resolved conversation
Escalations to humans cost nothing
Needs a Professional or Enterprise subscription

Put that next to a support rep. A human might handle 40 to 60 chats a day, and for simple repeat questions the cost gap is huge. Password resets, order status, and "where's my invoice" all belong with the agent.

It struggles with anything that needs judgment, refunds with exceptions, and angry customers who want a person. Also, the system decides what counts as "resolved," so watch your customer satisfaction scores and not only the resolution count.

Prospecting Agent: a drafter, not a closer

The Prospecting Agent watches for buying signals, researches accounts, and recommends leads with a drafted first message. Since April 14 it costs 100 credits, or $1, per lead it recommends for outreach, charged once per lead. You get a 28-day free trial and can set daily and monthly caps.

Here's my honest read. Every review I found says the same thing: the drafts need a human eye on nearly every one. Sybill calls it a "drafter with initiative," which matches what I'd expect from any AI writing cold outreach without deep context.

Is $1 a lead worth it? That depends on what your rep's time costs. Say the agent saves 10 minutes of research per account and your rep costs $40 an hour. That's about $6.70 of time for $1. The value is real, but don't expect it to run your outbound on autopilot.

A $1 lead that saves ten minutes of research is a good deal. A $1 lead sent without review is a brand risk.

Breeze Intelligence: enrichment inside the CRM

Breeze Intelligence is HubSpot's data layer. It draws on a database of more than 200 million company and contact profiles and adds three things:

Enrichment: fills in firmographic and contact data on your records
Buyer intent: shows which accounts are researching you
Form shortening: pre-fills known fields so your forms can be shorter
Standard enrichment costs no credits with paid seats. Intent is metered at 10 credits per company per month **for buyer intent, and **10 credits per record for Smart Properties.

I scrape and enrich data for a living, so here's my view: enrichment inside the CRM is convenient, but coverage swings widely by region and industry. If your market sits outside the US and Western Europe, test it on 100 records before you trust it.

How HubSpot credits work, and where costs creep in

Credits are the currency behind all of this. Here's the structure, per Vantagepoint's pricing breakdown:

Standard rate: $10 per 1,000 credits (about $9 with an annual commitment)
Starter includes 500 credits a month
Professional includes 3,000
Enterprise includes 5,000
Credits expire monthly and don't roll over
Run the numbers on Professional. 3,000 credits covers about 60 resolved chats or 30 prospecting leads. That's not much. A busy support inbox can burn through it in a day or two, and then you're buying more.

That's simply how usage pricing works, but it means your Breeze bill will rise and fall with your volume, so set caps from day one.

What I like and what bugs me

What I like:

Outcome pricing is the most honest AI pricing model I've seen from a major vendor
A free Breeze Assistant lowers the barrier for small teams
Everything sits in one CRM, so there's no data sync to manage
The Customer Agent's 65% resolution claim is strong if it holds for your use case

What bugs me:

The agents only see what's inside HubSpot. Sales calls, Slack threads, and other off-CRM context don't exist for them
HubSpot, not you, defines "resolved" and "qualified"
Professional's included credits run thin
The Prospecting Agent is pitched as autonomous but acts like an assistant

Who should turn on which part

Here's how I'd decide, based on where your team stands.

Any HubSpot user, free plan included: switch on Breeze Assistant. It costs nothing and helps right away.
Support teams with a solid knowledge base: trial the Customer Agent on one channel. Track CSAT next to resolution rate for 30 days.
Sales teams doing account-based outreach: trial the Prospecting Agent with a daily cap of 10 leads, and review every draft for the first two weeks.
Teams with messy data: clean it first. Every Breeze tool improves when records are tidy.
Teams outside HubSpot: Breeze alone is no reason to migrate. If you're already on HubSpot, though, it's a fine reason to stay.

A 30-day trial plan that won't burn your credits

If you test Breeze, run it as an experiment instead of a launch. This is the plan I'd follow.

Week 1: clean up and set a baseline. Choose one support channel and one sales segment. Record today's numbers: first response time, resolution time, CSAT, and how long reps spend researching an account. Without a baseline, every AI tool looks amazing.

Week 2: switch on the Customer Agent for one channel. Chat is usually the safest place to start. Set it to escalate anything involving refunds, billing disputes, or cancellations. At the end of the week, read 20 of its conversations yourself.

Week 3: add the Prospecting Agent with a cap. Ten leads a day, at most. A human reviews every draft before it goes out. Track how many drafts you send as-is, edit lightly, or throw away.

Week 4: do the math. Compare credits spent against time saved. If the Customer Agent resolved 300 chats for about $150 and CSAT held steady, that's a clear win. If most of the Prospecting Agent's drafts got binned, pause it.

The goal is to learn which part works for your team, not to prove Breeze works, and to turn off the rest before it quietly drains your credit balance.

The verdict

HubSpot Breeze AI shines when it stays quiet: the Assistant summarizing records, the Customer Agent answering the fifth "how do I reset my password" of the morning. That's where it earns its keep.

It's weakest where the marketing is loudest, which is the idea of an autonomous prospecting machine. That job still belongs to a human with an AI helper.

The outcome-based pricing is the real headline. If more vendors follow HubSpot and charge only when the AI does the work, that will change how every team buys AI. I hope they do.

Enjoyed this read? Follow me for more stories like it. You can also reach me on LinkedIn or Telegram @kawsarlog for updates on my work.

I take on freelance scraping projects, so if you need data collected, message me on LinkedIn or email hi@kawsarlog.com

NVIDIA Jetson Orin Nano 2: the small board that could power the next wave of robots

Md Kawsar — Mon, 12 Oct 2026 13:00:00 +0000

It packs 78 TOPS, 8GB of fast memory, and twice the inference of the $249 Orin Nano Super. Here's what that means for home robots, drones, and small builders.

Back in December 2024, NVIDIA made an unusual move. It took a $499 developer board, sped it up with a software update, and dropped the price to $249.

That board, the Jetson Orin Nano Super, became the go-to "first AI computer" for hobbyists, students, and small robotics teams. I've seen it in drone builds, inventory cameras, and even a friend's weekend plant-watering robot that mostly watered the floor.

Now there's a successor. On August 25, 2026, NVIDIA announced the Jetson Orin Nano 2, and its numbers could change what people mean by a "cheap robot."

Let's look at what's inside, what it might cost (and what's still unknown), and why I think this little board matters more than the flashy humanoid demos.

Why the entry level matters more than the flagship

This year, nearly every robotics headline has been about humanoids, with big companies, big budgets, and big videos.

Most real robots, though, are pretty boring: a floor cleaner, an inspection drone, a conveyor-belt camera that spots faulty parts. Whether these products succeed comes down to one question: can you fit enough AI into the device at a price customers will pay?

That's what an entry-level edge AI board is for. It's the brain inside the robot, running vision and language models right on the device so every frame doesn't have to go to the cloud.

The robot that leads the wave will be the cheapest one that's good enough, whatever the most impressive one can do.

Jetson Orin Nano 2 specs, side by side

This is how the new board stacks up against the Orin Nano Super, according to ServeTheHome's breakdown and NVIDIA's release.

AI performance: 78 TOPS (INT8 sparse) on Nano 2, up from 67 TOPS on the Super
CPU: 8 Arm Cortex-A78 cores, up from 6
GPU: 1,536 CUDA cores, up from 1,024
Memory: 8GB LPDDR5X-7500, up from LPDDR5-6400
Memory bandwidth: 120 GB/s, up from 102 GB/s
Power: 15W to 40W modes

Peak compute only rises about 16% on paper. So where does NVIDIA's "2x inference performance" claim come from?

Here's the interesting part. The Nano 2 uses new silicon rather than a relabeled Orin chip: Cortex-A78 cores alongside an Ampere GPU with architecture and microarchitecture changes. NVIDIA says those changes make real workloads run twice as fast, and that in 15W mode at the same performance, the new board draws 40% less power.

For a battery-powered drone or home robot, the power figure could matter more than the speed figure. Lower power means smaller batteries, less heat, and longer runtime.

What about the price?

I want to be clear on this, since plenty of posts are already calling it "the $250 AI computer."

NVIDIA hasn't announced Jetson Orin Nano 2 pricing yet. The module and developer kit should arrive in the first half of 2027.

What we do know is where it's positioned. NVIDIA says the board will "redefine entry-level edge AI," and it fills the same slot the $249 Orin Nano Super holds today. It's reasonable to expect a price in that range. Until NVIDIA confirms it, though, treat "$250" as an educated guess.

Same tier, same size, twice the useful work. If the price stays close to today's, the numbers for small robotics products look very different.

Which models can it actually run?

Here's where it gets practical. NVIDIA says the board supports its own Cosmos world models and Nemotron models, open models like Gemma 4 and Qwen 3, and lots of other open-source language and vision-language models.

Put simply, a small device running it can:

Look around a room with a camera and name what it sees
Follow a spoken command like "clean under the table"
Map out a route and respond when something moves
Handle all of it without an internet connection
That last one is an underrated win. Running models locally means faster responses, no cloud bill for every request, and better privacy. Nobody wants a home robot streaming their living room to a server.

Who's already building on it

NVIDIA named four early adopters, and together they cover a good slice of the "robot wave":

Cognex, for industrial machine vision
Doosan Bobcat, for construction and compact equipment
Matic Robotics, for home cleaning robots
Wing, Alphabet's drone delivery company
Matic's CEO, Navneet Dalal, summed up the home robot challenge nicely: robots have to understand people, map spaces accurately, and clean without help. That's three different AI tasks on one small board.

A long list of hardware partners, such as Advantech, AAEON, Seeed Studio, and Connect Tech, will also make carrier boards and cases for the module. That ecosystem is what turns a chip into products people can actually buy. NVIDIA says more than 3 million developers already use its robotics stack.

What this means if you're not a robotics company

My world is scraping and automation, so I'm looking at hardware from the outside. Here's how I see the Jetson Orin Nano 2.

Edge AI is becoming a normal line item. A few years back, adding a vision model to a warehouse camera meant a custom project. Soon it'll just be an option on the spec sheet, like choosing storage size.

Data work moves to the edge too. Devices that run models locally can also filter, label, and summarize data before it leaves the device. That changes how pipelines are built, with less raw video uploaded and more structured events.

For those of us who build data pipelines, that's a big change. Rather than gathering hours of footage to process in the cloud later, the device sends a clean stream of events: "pallet moved at 14:02", "shelf 4 empty", "person entered zone B". You get smaller payloads, cheaper storage, and data you can use as soon as it lands.

Prototypes get cheaper. If your small business idea needs "a camera that notices X," it gets easier with each generation. Think of a shop counting visitors, a farm spotting sick plants, or a lab keeping track of sample trays.

The catches worth knowing about

I like this board. Still, a few things are getting overlooked in all the excitement.

8GB is still 8GB. The memory is quicker, but the amount hasn't changed. That limits how big a language model you can run comfortably. Small and quantized models work well. Anything that needs a lot of room for long context will feel cramped. For many robots that's fine, because fast perception matters more to them than long conversations.

"2x" is NVIDIA's number. Actual speedups depend on your model, its precision, and how well your code uses the GPU. I'd wait for independent benchmarks on the exact models you plan to run before building a business case on it.

The top power mode isn't the efficient one. The 40% savings figure applies at matched performance in 15W mode. Run it at 40W for peak speed and battery life heads in the opposite direction. Choose the mode that suits your product, whatever the slide highlights.

The dev kit isn't the product. Most commercial robots mount the module on a custom or partner carrier board. Plan a budget for that engineering, plus cases, cooling, and certification. The chip is the cheapest piece.

A $250-class brain won't give you a $250 robot. What it does is make the brain one of the cheaper parts.
None of these are dealbreakers. They're simply the gap between a spec sheet and a finished product, and I'd rather you hear about them from me than from a surprised customer.

Should you wait or buy the current board?

If I were starting a project today, this is what I'd want to know. Here's my honest view.

If you're learning or prototyping now, get the Orin Nano Super. You can buy it now for $249, and everything you learn carries over, since NVIDIA's software stack stays the same.
If you're designing a product that ships in late 2027, design around the Nano 2. Because the form factor is unchanged, your mechanical design can stay as it is.
If battery life is your main constraint, the 40% power saving by itself justifies the wait.
If you need to ship something this year, skip waiting for hardware that's months away and build on what's available.

The robot wave is a price story

I keep thinking about that 2024 price cut. NVIDIA took the same chip, made it cheaper and faster, and brought a huge number of newcomers into edge AI.

The Jetson Orin Nano 2 feels like the next chapter of that story: twice the useful AI, lower power, the same size, and, if NVIDIA keeps the price, the same entry-level spot.

It may sound like a quiet headline. Yet the robots that make it into homes, farms, and warehouses will be the practical ones. They'll be the ones someone could afford to build.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

33 AI pricing and usage-limit changes in Q3 2026, collected in one free sheet

Md Kawsar — Sun, 11 Oct 2026 15:00:00 +0000

From July through September I logged 33 changes, and 22 landed in September alone. API prices mostly fell while subscription limits mostly tightened. Below are the patterns, the changes worth checking first, and a free sheet to match against your own stack.

If you pay for AI tools, your bill shifted this quarter, whether or not you spotted it.

From July 1 to September 30, 2026, I recorded 33 pricing and usage-limit changes across the AI vendors I follow, and 22 of them hit in September. Some were straightforward price cuts. Some were limit cuts dressed up as increases. A few were so quiet you'd only notice when a workflow died halfway through a task.

I covered several as they happened, including Three vendors restructuring usage limits in two weeks. This post gathers the full quarter in one place, along with a free sheet you can filter down to the tools you really use.

The whole quarter, summed up

API prices fell. Subscription limits tightened.

The sheet breaks down like this:

9 changes lowered costs, nearly all of them API or model prices
5 raised prices, mostly GPU compute and creative tool plans
5 cut access, mostly subscription limits and free plans
4 restructured how usage gets counted, which tends to make costs harder to compare
The remainder were new model launches, temporary boosts, mixed changes, and one retirement
Builders on APIs mostly got good news in Q3. Anyone on a flat-rate subscription got a reminder that promotional limits don't last.

Cheap AI has never been cheaper. Generous AI plans have never been harder to predict.

The changes to check first

Of the 33, these are the ones I'd look at first if you use the tools involved.

Claude Code: a +25% that's actually a 17% cut

Anthropic ran a +50% weekly limit promotion on Claude Code across the summer and extended it on July 19. On September 14, a permanent +25% baseline took its place. Digital Applied's limit ledger points out that this is a real gain over the original limits, yet a 17% reduction from what people had during the promo. Both descriptions are accurate, and which one you feel depends on when you signed up.

GitHub Copilot: promotional credits expire

On September 1, Copilot Business and Enterprise lost their promotional included credits. Monthly credits fell from 3,000 to 1,900 **on Business and from **7,000 to 3,900 on Enterprise, a cut of 37% to 44%. If your team leans on premium models in Copilot, open your usage reports now.

ChatGPT's five-hour limit returns

On July 12, OpenAI temporarily lifted the rolling five-hour limit for Plus, Business, and Pro users. Its return was announced for July 30, then pushed back. It finally came back on August 26, and the $100 and $200 Pro plans were exempt.

OpenAI's steep API cuts

On July 30, GPT-5.6 Luna fell 80% to $0.20 input and $1.20 output per million tokens. GPT-5.6 Terra fell 20% to $2 and $12, according to CloudZero's OpenAI pricing guide. For high-volume jobs such as classification or extraction, Luna at that price shifts the math on plenty of projects. Bulk parsing of scraped pages, the kind of work I do, is exactly what it suits.

Claude Opus 5.5 arrives cheaper than Opus 5

Claude Opus 5.5 launched at $4 input and $20 output per million tokens, under Opus 5's $5 and $25. Google's Vertex AI listed it on September 23, and Cursor swapped Opus 5 for Opus 5.5 at the lower rate on September 28. In late August Anthropic also cancelled a planned Sonnet 5 price increase, so it stays at $2 and $10.

Free plans are shrinking

Two quieter moves deserve a mention.** Replit removed its free Starter plan** on September 11, which makes the $20 Core plan the new entry point. And Pika's free plan now includes zero monthly credits. For some creative and coding tools, the "free forever" era is closing.

GPU compute got pricier

Renting GPUs now costs more. RunPod raised Secure Cloud rates, with the H100 SXM at $3.49 an hour. Nebius announced an H100 increase from $3.85 to $4.50 an hour, effective October 1. Fireworks AI lifted its US-only serverless premium from 10% to 50%.

Other changes worth a glance

These missed the top list, but one of them could be the one that hits your stack:

Apify cut its Starter plan from $29 to $19 a month on September 11. Good news if you run scrapers there, as I do.
Klaviyo swapped credit tiers for a $20 monthly plan with $10 top-ups on its AI Composer.
Suno limited commercial-use rights to your monthly download allotment.
Moonshot AI retired Kimi K2.5 and the entire moonshot-v1 API family on August 31. Anything pointed at those endpoints is now broken.
Beautiful.ai raised Pro from $12 to $14.50 a month on annual billing, and Hedra raised Basic from $15 to $20.
Zhipu AI added a Team tier to its GLM Coding Plan and set Individual Pro at $80 a month.
OpenAI launched GPT-6 Astra at $10 input and $50 output per million tokens, and Google launched Gemini 3.8 Flash at $0.75 and $3.75.
Each is small alone. Taken together, they show how often the ground shifts under anyone building with AI.

Three patterns behind the numbers

1. Promotions settle below their peak

Digital Applied's ledger found a steady pattern across 2026: temporary boosts get extended a few times, then land below the promoted level but above the original baseline. Claude Code and GitHub Copilot both followed it this quarter. Expect more of the same.

2. "Increased limits" often means "we changed how we count"

On August 24, Cursor told subscribers it had raised included usage but never published the size of the increase, and it moved Auto billing from a flat rate to per-model rates. Z.ai switched its coding plan from prompt counts to token credits. Atlassian re-metered automation as "steps." None of this is automatically bad. Still, once the unit changes, comparing old costs to new ones gets difficult, and that usually works in the vendor's favor.

3. Competition is driving API prices down quickly

Look at the evidence: Luna's 80% cut, Opus 5.5 priced under Opus 5, Perplexity's Fast Search API at 80% below its standard rate, cheaper embeddings from Voyage, and a 69% cut on Together AI's large-model fine-tuning. If you haven't re-checked your API model choices since spring, you're likely overpaying.

Announced and enforced are two different things. Check the date a change takes effect, not the date it was posted.

How to use the free sheet

The Sheet lists all 33 changes with the date, vendor, product, type of change, what changed, the direction for you, and a source link. Here's how to get value from it in ten minutes:

Filter by vendor down to the tools you pay for.
Enter "Y" in the "Affects me?" column for anything touching your stack.
Note one action per row: switch models, change plan, or ignore.
Open the Summary tab, which counts changes by direction and month, plus how many you flagged as affecting you.
Confirm on the vendor's page before acting, because prices move fast.

What I'd do before Q4 budgets

If you manage AI spend for yourself or a team, here's a short action list:

Re-price your API workloads. Run your top three tasks on the newly cheaper models. Luna, Terra, and Opus 5.5 all changed the options.
Check your coding tool limits. Claude Code, Copilot, and Cursor all changed this quarter. Look at real usage rather than what the plan page promised in June.
Leave disappearing free plans before they cause trouble, especially if a client workflow depends on one.
Lock in GPU capacity if you rent compute and your provider's prices are climbing.
Schedule a 15-minute monthly review of a pricing tracker such as UsagePricing's activity log.

A note on sources

The sheet draws on public trackers, chiefly Digital Applied's coding-plan limit ledger, UsagePricing's activity log, and CloudZero's OpenAI pricing guide, along with vendor coverage. It covers the changes that touch the tools most builders, freelancers, and small teams use, which means some vendors are missing. Each row links to its source, so you can verify anything that matters to you.

If I missed a change that hit your stack, tell me in the comments and I'll add it to the next quarterly sheet.

Ten minutes before Q4

Q3 2026 produced 33 AI pricing and usage-limit changes worth knowing. The short version: API prices fell, subscription limits tightened, and several "increases" were really cuts from a promotional peak.

You don't have to follow all of it. You just need to know which rows touch your stack. Download the free sheet, filter it to your tools, and give it ten minutes before you plan next quarter's budget.

Enjoyed this read? Follow me for more stories like it. You can also reach me on LinkedIn or Telegram @kawsarlog for updates on my work.

I take on freelance scraping projects, so if you need data collected, message me on LinkedIn or email hi@kawsarlog.com

Free vs paid AI tiers in 2026: what you really get for $0 from each vendor

Md Kawsar — Sat, 10 Oct 2026 13:00:00 +0000

ChatGPT's free plan now shows ads in some markets. Claude's free plan has no Claude Code. Google AI Studio's free tier may use your content to improve its products. Here's what $0 buys on each, and when paying makes sense.

Whenever someone asks me which AI subscription to buy, I say something that catches them off guard: maybe none, at least not yet.

In 2026 the free tiers are strong. You can talk to frontier-level models, search the web, upload files, and sometimes even get API access without spending anything. But free always costs something. Sometimes it's ads, sometimes it's your data, and sometimes it's a usage cap that kicks in at the worst possible moment.

A few of these tools are on my paid list because my scraping and automation work relies on them. I still follow the free tiers closely, since they're what I point clients, friends, and beginners toward. Below is my straight breakdown of free vs paid AI tiers from the major vendors.

The short version

Short on time? Read this part:

Best free chat for most people: ChatGPT Free or Gemini, depending on how much of your day happens inside Google's apps.
Best free option for developers: Google AI Studio with the Gemini API free tier.
Best reason to pay $20: you run into limits every week, or you need coding agents, deep research, or long-document support.
Biggest hidden cost of free: your data may be used to train models, and some free plans have started showing ads.
Free AI stopped being a trial. It's a real product now, and like most free products it charges you in data, attention, or limits.

ChatGPT: generous, but with ads now

Here's OpenAI's lineup, as laid out in GeoToolbox's September 2026 breakdown

What Free includes

GPT-5.6 Luna as the default model
Text chats without a hard cap, which arrived in mid-August
Image generation and file uploads
Five lightweight research runs per month

The catch

In some markets, Free now shows ads. You can switch to an ad-free mode, but it costs you some daily messages. Your conversations may also be used for training unless you opt out in settings.

What paying adds

Go ($8/month): raised file and image limits, the same default model, and still ad-supported where ads exist
Plus ($20/month): GPT-5.6 Sol with adjustable effort, Deep Research, Codex, Agent Mode, Projects, and no ads
Pro ($100 or $200/month): about 5x or 20x the Plus usage, a larger context window, and more Codex

Claude: strong free chat, but no coding agent
These details come from Vantagepoint's September 2026 guide and Anthropic's pricing page

What Free includes

Chat on web, mobile, and desktop
Web search and file creation
Connectors to other apps

The catch

Free carries the lowest usage limits of any Claude plan and no Claude Code. For the coding agent, you'll need Pro.

What paying adds

Pro ($20/month, or roughly $17 per month on annual billing): Claude Code, unlimited projects, the Research tool, and higher limits
Max ($100 or $200/month): 5x or 20x Pro usage
Team ($20 to $125 per seat): admin controls, SSO, and no training on your content by default
Worth knowing: Claude Code limits moved several times this year. A temporary +50% weekly boost eventually settled into a permanent +25% baseline in mid-September. If you pay mainly for coding, look at your actual weekly limit instead of assuming.

Gemini app: free across Google's products

What Free includes

According to this 2026 comparison the free Gemini tier runs on the Gemini 3.x family with usage limits. It's built into Search, Gmail, and Docs, and it includes limited image and video generation.

The catch

You get a smaller context window than paid plans, and heavy users will run into daily caps.

What paying adds

Google AI Pro ($19.99/month): 1M+ token context, most caps lifted, more image and video generation, plus Google One storage
Google AI Ultra: top limits and early access to new features, at a premium price
Already paying for Google One storage? Then AI Pro may cost less than it looks, because that storage comes bundled in.

Google AI Studio: the developer's free lunch

This is the free tier I recommend most often to developers, freelancers, and anyone with a side project.

What Free includes

Google AI Studio lets you try prompts across Gemini models, upload images, PDFs, video, and audio, build simple agents, and create an API key. None of it needs a credit card. The Gemini API free tier has rate limits, but it doesn't expire.

The catch

Here's the big one. On the free tier, Google may use your content to improve its products. Paid tiers don't. So keep client data, private code, and anything confidential out of it. Rate limits also differ by model and shift often, so check the official limits page before you build a plan around them.

What paying adds

Pay-as-you-go API pricing. Take Gemini 3.8 Flash: it costs $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, with 50% off batch jobs and up to 90% savings on cached context.

If it's free and it touches your data, assume the data is part of the price. Read the terms before you paste.

Free API credits from other providers

Building something? A few other providers give you a free place to start, per Admix's 2026 guide:

OpenAI and Anthropic: small signup credits that expire
DeepSeek and Mistral: rate-limited free tiers meant for development
Groq: a free tier for fast open-model inference
Cloudflare Workers AI: a free daily allowance
Startup programs: the big clouds and labs offer far larger credits if you qualify
A practical approach is to stack them. Prototype on the free tiers, then move the one workload that proves itself onto a paid plan.

The best $0 setup for three kinds of people

If I were assembling a fully free AI toolkit today, this is what I'd choose.

The student or career switcher

ChatGPT Free for everyday questions and explanations, Gemini inside Google Docs for writing help, and Claude Free as a second opinion on anything that matters, like a cover letter or a hard concept. That's three free tools with three different voices, which helps when you're learning.

The freelancer or solo business owner

Pick one free chat app you enjoy for drafting and thinking. Use Google AI Studio to test prompts on real tasks, but feed it made-up or public sample data only. Add a free email or document tool with built-in AI for small jobs. Once client work depends on it, pay for one plan that doesn't train on your data.

The developer building a side project

Prototype on the Gemini API free tier. Handle quick classification or extraction steps with a free open-model tier like Groq. Use signup credits from OpenAI or Anthropic to compare quality on your toughest prompt. And log every request, so you know the real cost the day you move to paid.

Not one of these setups costs a cent, and each is good enough to learn on, test ideas, and even ship something small. That's remarkable, and it wasn't true two years ago.

When to upgrade (and when to hold off)

Here's the rule I give people. Upgrade if at least one of these is true:

You hit limits more than once a week. If you use AI daily, waiting for resets costs more than $20 in lost time.
You need a specific paid feature. Claude Code, Codex, Deep Research, Agent Mode, or long context for large documents.
You handle anyone else's data. Client work belongs on a plan that doesn't train on your data, ideally a team plan with admin controls.
Ads or interruptions wreck your focus. Some people don't mind them. Plenty do.
And don't upgrade just because a new model came out. Free tiers usually pick up newer models within a few weeks.

Pay for one plan, not three

People often make the same mistake. They pay for ChatGPT Plus, Claude Pro, and Google AI Pro at once, then use one of them about 80% of the time. That's $60 a month for maybe $25 of value.

Try this instead:

Run the free tiers of all three for a month.
Notice which one you open first when something matters.
Pay for that one.
Keep the others on free for second opinions.
For coding-heavy work, I'd favor whichever plan includes the coding agent you like best. For research and writing, test each on your real tasks for a week. And if you live in Google Workspace, AI Pro's bundled storage makes the decision easy.

Start free, pay for one

The question of free vs paid AI tiers in 2026 doesn't come down to good versus bad. The free plans are strong. What separates them from paid is limits, power features, privacy, and now ads.

My advice: start free, keep client data off the free tiers, and upgrade the one tool you depend on once limits start eating your time.

Check back every few months, too. These plans shift constantly, and today's best deal probably won't hold up by January.

Enjoyed this read? Follow me for more stories like it. You can also reach me on LinkedIn or Telegram @kawsarlog for updates on my work.

I take on freelance scraping projects, so if you need data collected, message me on LinkedIn or email hi@kawsarlog.com

Smartlead vs Instantly: which cold email tool fits how you actually work

Md Kawsar — Fri, 09 Oct 2026 13:00:00 +0000

Their mid-tier plans are only about $3 apart. What really separates them is the audience: in-house teams that want lead data built in, or agencies that need control. Here's everything side by side.

Every cold email agency I know has this debate at least once a quarter. One person says, "Let's move everything to Instantly." Another says, "No, Smartlead works better for clients." And then out come the spreadsheets.

I see that debate from both ends. I scrape B2B leads for clients and I run an email verification platform, so I watch the data go in and the bounces come back. Smartlead vs Instantly is a question I hear all the time.

Here's the answer I really give people, with 2026 pricing, limits, and the details you won't find on either landing page.

The short answer first

Short on time? Here's the quick version:

Instantly suits in-house teams that want a polished tool with a lead database, placement testing, and a light CRM in the box.
Smartlead suits agencies juggling lots of client accounts that need separate workspaces per client, white-label dashboards, and solid API control.
The rest of this post explains why.

The prices are nearly the same. The workflows are very different. Choose the tool that fits how your team really runs campaigns.

2026 pricing, side by side

These are this year's published plans, based on GigRadar's 2026 comparison. Check each vendor's pricing page before you buy, since both change their plans often.

Smartlead

Basic: $39/month, 2,000 contacts, 6,000 emails/month
Pro: $94/month, 30,000 contacts, 150,000 emails/month
Unlimited Smart: $174/month
Unlimited Prime: $379/month

Instantly

Growth: $47/month, 1,000 contacts, 5,000 emails/month
Hypergrowth: $97/month, 25,000 contacts, 100,000 emails/month
Light Speed: $358/month

On the most popular middle plans, Smartlead Pro ($94) vs Instantly Hypergrowth ($97), the gap is just $3. Smartlead gives you slightly more contacts and sends, while Instantly packs in more features.

Both include unlimited connected email accounts and unlimited warmup on every plan. A few years ago that set them apart. Now it's standard.

The hidden costs

This is where the sticker price can mislead you.

Smartlead's add-ons: inbox placement testing costs extra, somewhere around $49 to $599 depending on volume, according to Artisan's breakdown. White-label client dashboards cost about $29 per client each month. There's no lead database built in, so you need your own data source.

Instantly's bundles: placement testing and a light CRM come included. There's also SuperSearch, a lead database with more than 450 million records. If you'd otherwise pay for Apollo or something like it, that saves you real money.

That means an in-house team that needs both leads and sending can spend less overall on Instantly, even with its higher base price. An agency that already has its own data keeps Smartlead's low price low.

Deliverability: the part that actually matters

Here's something people don't love hearing. Your tool matters less than your infrastructure. Your domains, DNS records, list quality, and sending volume decide whether emails reach the inbox. The platform is mostly the steering wheel.

Still, the two tools do differ:

Both use shared warmup networks. Some Reddit users say their open rates dropped suddenly because of warmup pool quality on Instantly, and GigRadar mentions it too. I'd call that anecdotal, but it's worth watching.
Instantly comes with placement testing, so you can see whether you're hitting inbox or spam before a campaign scales up.
Smartlead's mailbox rotation across unlimited accounts spreads volume thinly and evenly, which is exactly what agencies want.
And the basics still apply whichever tool you choose. Since 2024,Google's sender guidelines have required authentication and low spam complaint rates, and Microsoft tightened its rules for bulk senders in 2025. If SPF, DKIM, and DMARC aren't set up yet, close this post and fix that first.

Send a bad list through a great tool and it still bounces. Verify before every send.

Inbox management and replies

Replies are where deals get made, so this part deserves more attention than it gets.

Smartlead's Master Inbox gathers replies from every account in one place, with custom categories, webhooks, and the option to assign replies to team members. For an agency with several SDRs covering different clients, that assignment feature is a big help.

Instantly's Unibox handles the same job with a cleaner design and AI that labels replies automatically, such as interested, not interested, or out of office. For one in-house team, it's quicker to work in every day.

Agency features: Smartlead's home turf

If you run an agency, this is the section that settles it.

Smartlead gives every client its own separate workspace with isolated API tokens
White-label dashboards show clients their results under your brand
The API is solid enough to build custom workflows, reports, and integrations
Instantly offers workspaces as well, but reviewers say they feel more like saved campaign folders than truly separate client accounts. With two clients, that works fine. With twenty, Smartlead's setup saves you a lot of trouble.

What real volume costs: a worked example

Let's put real numbers on it. Imagine you want to send 30,000 cold emails a month, a typical goal for a small agency client or a committed in-house SDR team.

Inboxes first. If each inbox sends about 40 cold emails a day over roughly 22 sending days, it handles around 880 emails a month. That means you need about 35 inboxes. Spread them over something like 12 domains with 3 inboxes each, and you have a normal, safe setup.

Platform cost. Smartlead Pro ($94) and Instantly Hypergrowth ($97) both handle 30,000 sends a month easily. Since accounts are unlimited, connecting 35 inboxes costs nothing extra on either one.

The real bill. The platform is the cheapest piece. Domains and mailboxes cost a lot more. Google Workspace or Microsoft 365 seats run from a few dollars up to about $7 each, so 35 inboxes can cost $100 to $250 a month, before domain renewals. On top of that come lead data and verification.

So the total monthly cost for this setup often lands between $400 and $700, and the gap between Smartlead and Instantly barely registers. That's why I keep telling people to choose based on workflow.

AI features: handy, but they shouldn't decide it

Both tools push AI hard these days. Instantly offers AI reply labeling in Unibox and AI help with writing campaigns. Smartlead offers AI sorting in its Master Inbox and personalization helpers.

Honestly, AI-written cold emails sent at scale start to sound alike, and prospects can tell. The AI I'd really use is the boring kind: sorting replies so your team gets to interested leads first. Both tools do that fairly well, so it shouldn't be the reason you switch.

Which one should you pick?

This is how I'd make the call.

You're a founder or small in-house sales team without a lead source yet: Instantly. Getting SuperSearch and sending on one bill is tough to beat.
You're an agency with 5+ clients: Smartlead. Its workspaces, white-label options, and API control were made for you.
You already pay for Apollo, Clay, or scraped data: Smartlead's lower base price is the better deal, because you won't need Instantly's database.
You're developer-heavy and want to automate everything: Smartlead's API gives you more flexibility.
You want the least setup friction: Instantly's interface is smoother for non-technical users.

If you're thinking about switching

Switching is easier than most people expect, because the valuable parts sit outside the tool. Your domains, inboxes, and warmup history belong to your mailboxes, so neither Smartlead nor Instantly keeps them.

Here's how a clean switch works. Connect your current inboxes to the new platform and let warmup run there for a week while your old campaigns wrap up. Export your lead lists with their status so you don't email anyone who already replied or unsubscribed. Rebuild your best sequence first and leave the rest for later. Then keep the old account for one more billing cycle so you can look back at reply history.

The most common mistake I see is switching tools to fix deliverability. If your emails go to spam, the same domains and the same list on a new platform will go to spam as well. Fix the real cause first.

My cold email infrastructure checklist

No matter which tool you pick, I'd have this setup in place before the first campaign goes out.

Separate sending domains from your main domain. Keep your company inbox out of harm's way.
SPF, DKIM, and DMARC configured on every sending domain.
Warm up for at least two weeks before sending real campaigns.
Cap each inbox at a sane daily volume. Lots of teams keep it to around 30 to 50 cold emails per inbox each day.
Verify every list. Remove invalid and risky addresses, plus any catch-all addresses you can't confirm.
Test placement before you scale up a campaign.
Keep an easy opt-out in every sequence. It protects your sender reputation and shows respect for the people you email.

The verdict

The Smartlead vs Instantly debate often gets treated like a boxing match. In reality, the two have grown into tools for two different jobs.

Instantly is the all-in-one option for teams that want data and sending in one tidy package. Smartlead is the control panel for agencies managing lots of clients.

Choose the one that matches your workflow, then put your real effort into domains, lists, and copy. That's what decides whether cold email works. The tool just sends it.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

$150,000 per song: what the AI music lawsuit wave means for every AI company

Md Kawsar — Thu, 08 Oct 2026 15:00:00 +0000

Sony and Universal have pinned a theoretical $9 billion price on Suno and more than $4 billion on Udio. Here's where those numbers come from, why nobody expects them to be paid, and what AI builders should take away.

If you run an AI company, this number should get your attention: $9,153,900,000.

You get it by multiplying 61,026 sound recordings by **$150,000 **each, which is the highest statutory damage US law allows for willful copyright infringement. That's the theoretical maximum Sony Music and Universal Music Group have been pointing at Suno

No serious person expects Suno to pay nine billion dollars. The number was never supposed to be realistic. Its job is to be scary, and it's doing that job well.

I'm not a musician. My work is data pipelines and scrapers. Still, the AI music lawsuit fight is the best preview I've found of how training-data disputes will go for the rest of us. Let's walk through it.

How we got here, fast

Here's a quick summary of a long and messy timeline:

June 2024: The RIAA sues Suno and Udio for the major labels.
October 2025: Independent artists begin filing class actions against Suno and Udio.Later suits go after Mureka and Google's Lyria 3.
Late 2025: Warner Music settles with Suno, and Udio signs licensing deals with Universal and Warner.
May 21, 2026: Sony and UMG ask to add 61,026 recordings to their Suno case. They identified them with the Audible Magic fingerprinting service.
June 29, 2026: A judge won't let Sony add tens of thousands of new works to its Udio case.
July 20, 2026: Sony files a brand new lawsuit against Udio that covers more than 30,000 recordings.
September 2026: Sony and UMG sue Suno a second time, now aiming at its new v6 model.
So this is a whole wave of lawsuits, and each new one is bigger than the one before.

Where the $150,000 number comes from

Under US copyright law, a rights holder can pick statutory damages rather than proving what they actually lost. The normal range is $750 to $30,000 per work. When the infringement is willful, a court can award up to $150,000 per work.

Everything hinges on that per-work rule. How much a label lost on any one song doesn't matter. The math simply counts songs.

Here's how that counting adds up:

Original Suno complaint: 560 works, about $84 million at the max
Expanded Suno list: 61,026 works, about $9.15 billion at the max
Sony's second Udio suit: 30,000+ works, about $4.5 billion at the max

With statutory damages, your training dataset becomes a bill, and every file is a line on it.

The new twist: "model laundering"

To me, the September Suno suit is the most interesting one, because it targets a new model.

Suno says v6 was trained on licensed material, community interactions, and what its team learned. The labels say there's more to it. Their complaint claims Suno used knowledge distillation, meaning v6 learned from the outputs of older models that had been trained on unlicensed tracks. Variety quotes the complaint calling v6 "the fruit of the same poisoned tree."
If the court accepts that argument, it changes a lot. Retraining alone wouldn't wipe a model's history clean. The taint follows the lineage.

A huge number of AI products today come from distilling a large model into a smaller one. If courts decide distilled models inherit their parent's legal problems, plenty of companies will need to take a close look at their family tree.

Why the labels won't actually collect $9 billion

Headlines love these figures, so I want to be clear: the maximums are upper limits, and nobody should read them as forecasts.

Here's why they almost never get paid in full:

Willfulness has to be proven, and courts usually award much less than $150,000 per work
Judges resist cases that keep ballooning, as the Udio ruling showed
Fair use is still unsettled, and results vary from case to case
Most of these fights wrap up with licensing deals rather than damage awards
That last point is the pattern to watch. Warner settled and signed a license. UMG and Warner licensed Udio. Suno now has deals with Warner, BMG, and Believe. These lawsuits act less like a path to a verdict and more like a negotiation backed by a very big stick.

The artists who aren't in the room

The label cases grab the headlines, but the independent artist cases could end up mattering more.

The Class actions filed since October 2025 speak for songwriters, session musicians, producers, vocalists, and small labels. Their argument is straightforward: the systems learned from full tracks, vocal stems, and lyric sheets, and now produce music that competes with the very people who made the originals.

This matters because a major label's licensing deal only covers that label's catalog. A session player who appeared on hundreds of recordings, or an indie artist who released music alone, sits outside that deal. Warner or Universal getting a seat at the table doesn't give them one.

Even if every major label signs on, a long tail of rights holders can still sue. For AI companies, "we signed the big three" makes the risk smaller, but some of it is still there.

It also leads to a tougher question that nobody has solved yet: how do you fairly pay thousands of small creators for a model trained on millions of files? Per-stream royalties took years to work out, and per-training-example royalties haven't even begun.

What this means for AI companies outside music

This is the part that matters most to me. Music is only the first test. The same reasoning covers images, code, news articles, product data, and anything else collected at scale.

Your exposure is your file count

Because damages are counted per work, a company trained on 10 million copyrighted items has an exposure on paper that sounds ridiculous. That figure can still surface in investor due diligence, acquisition talks, and insurance quotes, even if it never reaches a courtroom.

Provenance is becoming a product feature

Saying "here's exactly where our training data came from" is becoming something you can sell. Udio and Suno both advertise their licensing deals now, and people buying AI tools are starting to ask about it.

Fingerprinting makes hiding hard

The labels used Audible Magic to find 61,026 of their tracks. Similar tools exist for images, text, and code. Betting that nobody will check your dataset is a bad bet.

People used to ask "is this legal?" Now they ask "can you prove where it came from?"

A practical checklist if you train or fine-tune anything

Scraping data for clients is my everyday work, so I've thought about this a lot. This is what I'd tell any small AI team.

Keep a data manifest. Record the source, the collection date, the license or terms, and how you used it. It's a dull spreadsheet that pays off later.
Separate training data from reference data. Collecting public prices for analysis carries a very different risk from feeding creative works into a generator.
Check the terms of every source. Just because something is public doesn't mean you can train on it.
Track your model lineage. If you distill from another model, note which one it was and what it was trained on, as best you know.
Prefer licensed or open datasets for anything that produces output competing with the original creators.
Talk to a lawyer before launch, instead of after a demand letter arrives. I'm not a lawyer, and this isn't legal advice.

Dates worth putting in your calendar

If you'd like to keep up without reading every filing, these are the key dates in the main Suno case, according to Music Business Worldwide:
September 30, 2026: fact discovery and depositions end. Any evidence about training data is now on the record.
April 9, 2027: dispositive motions are due. Each side will ask the judge to decide big questions, such as fair use, without a full trial.
Keep an eye on whether the court allows the 61,026-work expansion. The Udio judge rejected a similar request. If the Suno judge does the same, expect more brand-new lawsuits instead of bigger existing ones, just like Sony did with Udio.

The honest read

My feelings about this fight are mixed.

I understand why musicians are angry. If your work were used without permission to build a machine that competes with you, you'd be furious too. The independent artist class actions seem especially fair to me.

At the same time, "$9 billion" doesn't help anyone understand what's really happening. It's a number built to apply pressure. What actually happens, over and over, is a licensing deal where the biggest labels get paid and smaller creators wait to see if anything reaches them.

For AI builders, the lesson is to keep building and know your data the way you know your finances. In 2026, your dataset can turn into an itemized bill, and the people holding the receipts have gotten very good at counting.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

A Yemen weapons cell used Claude as its engineering team. Here's how Anthropic caught it

Md Kawsar — Wed, 07 Oct 2026 13:00:00 +0000

A small group leaned on Claude Code to write missile software. How it got caught says a lot about the way AI companies watch over their tools in 2026.

Somewhere in northern Yemen, a guided rocket failed its test. A few hours later, the people behind it were back in a chat window, asking an AI to explain what broke.

Of everything in Anthropic's September 2026 threat intelligence report,that's the detail I keep coming back to. I skimmed past the missile ranges and the geopolitics. What stayed with me was the reflex. Something failed, and their first move was to open Claude.

Claude is part of my daily work for scraping jobs and client automation. So when I read that a weapons cell "used Claude Code in place of human software engineers," two feelings hit me together. I felt a bit sick. I also really wanted to know how they were caught.

This post is about that second feeling. AI misuse enforcement has changed a lot without much noise, and if you build anything with AI tools, the new rules affect you as well.

What actually happened in Yemen

According to the report and coverage from Al Jazeera and Axios,operators in Houthi-held northern Yemen were working on three weapons programs: a guided rocket, a multi-stage ballistic missile, and another missile family that included a planned hypersonic glide vehicle.

They went further than a single chat. They set up a small org chart made of AI.

One Claude instance handled the coding
A second one did the research
A third checked the output, the way a senior engineer reviews code
Claude helped with flight-control code, simulations, and fixing guidance and navigation problems. When a guided rocket test failed, the team returned to ask for diagnostics. Anthropic says it has no evidence the group fielded a working weapon. Still, the group had built offline simulation tools too, so it was already becoming less dependent on Claude.

Putting together a three-person engineering team used to mean hiring. For this cell, it just meant writing prompts.

How they slipped past the first line of defense

Most headlines left this part out. Anthropic's safety systems did stop many of their requests. The cell worked around them with two simple tricks.

First, they hid the objective. No one wrote "help me build a missile." They asked about control loops, simulation math, and debugging, and each of those questions looks ordinary by itself.

Second, they split the project across many separate conversations. Every session looked innocent. The danger only became clear once you put them side by side.

Anyone who has scraped a site with per-session rate limits knows this trick. You spread the work out so no single request stands out. It's the same idea, used for a much darker purpose.

That's why filtering prompts on their own isn't enough anymore. A filter that reads one message at a time will miss a pattern that only appears across fifty.

What AI misuse enforcement looks like in 2026

After reading the whole report, I'd put the new approach in one sentence: enforcement moved from judging prompts to judging behavior.

Anthropic lays out several layers that work together.

Classifiers on individual requests. This is the original layer. It still stops a lot, and careful wording can still get around it.
Behavioral signatures across accounts. This layer watches account patterns, long-running workflows, reused templates, and sessions that look coordinated. The report says influence operators pasted the same Markdown "doctrine" files, almost word for word, into hundreds of sessions. Repetition like that works like a fingerprint.
Dedicated investigations. The Yemen case surfaced through internal investigations into weapons development, and people on Anthropic's team followed the leads by hand.
Outside tips and cross-platform data. Some tips arrive from partners or public reporting, and the team then checks them against activity on the platform.
Action after detection. Accounts are banned. The actor's behavior is turned into new automated detections, so the next similar operation gets flagged sooner. Indicators are passed on to government and industry partners.

The fifth layer interests me most. Each operation that gets caught teaches the detector something new. The system gets better by studying the people trying to beat it.

The Yemen case wasn't the scariest thing in the report

I went in assuming the missile story was the biggest news. After reading the rest, I'm less sure.

It covers a Chinese-speaking group, probably students plus someone interning at a security firm, running what Anthropic calls an autonomous "zero day exploit foundry." Swarms of agents took binaries apart in parallel and wrote exploits without human help.

There's also a money-driven crew that scanned 1.8 million Android APKs for leaked credentials and collected more than 2,100 Azure AD tokens in 34 hours. And a lone French-speaking hacktivist who built a full doxxing platform alone.

This is the line I can't shake:

"AI has collapsed the labor and tooling gap that used to separate state-sponsored operations from individual operators."
To me, that's the real headline. Sophistication no longer tells you who's behind an attack. One person with a strong agent setup can look like a small agency. The Yemen cell is simply the most dramatic example of that change.

Why a freelancer or small founder should care

Maybe you're thinking: I collect product data and build dashboards, so why should missile software matter to me?

It matters more than you'd expect. Three parts of this report affect everyday builders directly.

Your API keys are now loot

One Russian-speaking actor targeted AI vendors' evaluation sandboxes and took production API keys from about 30 AI companies in four days. According to the report, criminals now see AI keys the same way they see production credentials. A stolen key gives them something to sell, free compute, and cover, because every action traces back to your account instead of theirs.

Let that last point sink in. If someone takes your key and uses it for abuse, your account is the one that gets flagged.

Behavior-based enforcement can catch honest patterns too

When detection focuses on workflows and templates, heavy but honest automation can look odd. I run long agent sessions for data extraction and often reuse the same instruction files from job to job. That's ordinary for my work. On paper, though, it's the same kind of repetition detectors look for.

Honest users don't need to panic. It's still smart to keep your usage easy to explain, with clear project names, reasonable request rates, and a record of what each automation does.

"Split it across sessions" is not a clever workaround anymore

Some people try to slip past content policies by chopping a request into harmless-looking pieces. The Yemen case shows that this now gets investigated as a pattern. Legitimate work doesn't need tricks. And if the work isn't legitimate, the tricks are what draw attention.

A practical checklist for builders using AI tools

After reading the report, this is the checklist I'd give any freelancer or small team. Every item is easy.

Rotate API keys on a schedule. Do it every 60 to 90 days, and right away whenever a contractor leaves.
Never ship keys inside apps or public repos. The APK scanning case proves attackers dig through compiled apps. Route calls through a backend proxy instead.
Set hard spend limits on every AI provider account. A sudden jump in spending is often the first clue that a key leaked.
Separate keys per project and per client. Then a leak means revoking one key instead of all of them.
Keep a short note of what each automation does. If your account is ever reviewed, you'll have a clear answer ready.
Read the provider's usage policy once a year.It changes more often than most people realize. Here's Anthropic's if you've never looked.
Watch your own agents. Autonomous agents with web access and saved credentials have become targets. Keep logs of what they do.

What I'm still uneasy about

To be honest, I'm still not sure how I feel about all this.

On one side, the system did its job. A weapons cell was found, banned, and reported, and the detectors learned from it. That's a good outcome.

On the other side, the cell did get real help before anyone stopped it, and it still has simulation tools it can run offline. The report also says that when state-backed researchers were refused on a biology request, they just switched to a competing model. One company's enforcement only reaches that one company.

Then there's the privacy question people tend to avoid. Detection based on behavior means someone is studying behavior. That's great when it's aimed at bad actors. For everyone else, it's a trade-off we're accepting without much debate.

I don't have a tidy answer. I just think anyone building with AI should see this trade-off clearly and stop pretending it isn't there.
The takeaway

Most people will remember the Yemen story as "terrorists used AI for missiles." That's accurate, and it's frightening.

For most of us, though, the more useful lesson is a quieter one. AI misuse enforcement now watches patterns as well as prompts. Stolen keys become disguises. And small teams can work like large ones, for better and for worse.

So guard your API keys the way you guard passwords. Keep your automation boring and easy to explain. And next time a threat report comes out, look past the scary headline for a moment and study the method. That's where the lessons for your own business are.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com