DEV Community: Skila AI

I Ranked Every AI Image Model by Speed. The $0.01 One Crushed GPT Image 2.

Skila AI — Tue, 12 May 2026 07:19:14 +0000

I ranked every AI image generator on the May 2026 leaderboards by one number: seconds per image. Not Elo score. Not how pretty the output looks at 1080p. Just: how long does a user wait from prompt to pixels.

The result reordered everything I thought I knew about this category.

The fastest model in production right now is not from OpenAI, not from Google's flagship line, and not from Midjourney. It's Z-Image Turbo, an open-tier model that ships images in about a second for one cent each. Meanwhile GPT Image 2 — the model topping the quality Elo at 1338 — can take a full minute on a complex prompt. That's a 60x latency penalty for marginal quality gains most apps will never surface.

I'll walk the full ranking, the news that made today the moment to read it, and how to pick a model for the job you actually have.

The news anchor: why today, May 11-12 2026

Two things landed in the last 48 hours that make this an unusually clean snapshot.

First, OpenAI rolled out GPT-5.5 Instant to all ChatGPT users on May 11. Instant means a faster default tier — and OpenAI is pulling latency forward across the stack, including its image side. The bar for what counts as "fast" just moved.

Second, Google's Gemini Omni video model leaked ahead of Google I/O 2026. Nano Banana 2 (the codename for Gemini 3.1 Flash Image) is hitting peak API adoption right now as devs migrate ahead of Omni. If you're picking an image stack this week, you're picking it on top of a market that's about to get reshuffled again.

Speed numbers move every few weeks. The ones below are pulled from the published vendor leaderboards (llm-stats, Artificial Analysis, Atlas Cloud's 2026 benchmark) and read at standard tier unless noted.

The full ranking (May 2026)

1. Z-Image Turbo — ~1 second, $0.01/image

Cheapest and fastest on the board. llm-stats has it sitting at Elo 302, which puts it near the high-end pack for quality despite the cost. This is the model to default to for chat-UX scenarios where the user is staring at a spinner. If anything beats it on speed at this price tier I haven't found it.

2. Google Nano Banana 2 (Gemini 3.1 Flash Image) — 1-3 seconds standard, $0.067/image

The speed leader at API scale. "Standard" tier finishes in 1-3 seconds; flipping to the 'Pro' tier on the same family pushes you to 4-6 seconds but bumps fidelity. Google has been quietly winning the latency war here for two release cycles — this is the safe default for production apps that need consistent quality and speed, not just one or the other. Google AI Studio is the canonical UI if you want to test it without writing API code.

3. Seedream v5 Lite — ~2 seconds

The dark horse from ByteDance. v5 Lite is genuinely fast at high resolution — most competitors slow down by 2-3x at 2048x2048; Seedream barely flinches. If you've used Dreamina, you've already touched the Seedream stack — Dreamina is ByteDance's consumer frontend over the same models.

4. Imagen 4 Fast — ~3 seconds

Google's text-rendering specialist. If your prompts include real words inside the image (signage, labels, packaging), this is where to start. Slower than the top three but the text doesn't garble.

5. Flux 1.1 Pro — ~6 seconds

Black Forest Labs' photorealism leader. 6 seconds is the cost of looking like a photograph. Worth it for hero shots, ad creative, anything where the audience is supposed to forget it's synthetic.

6. Midjourney v7 — multi-second turbo / 10-30s standard

Still the artistic ceiling, still the slowest mainstream option. Midjourney v7 in turbo mode is acceptable; in standard mode it's a non-starter for batch generation. Workflow: use it for the one frame that has to look like an oil painting, not the gallery wall.

7. GPT Image 2 (standard) — ~15 seconds simple / 40-60s complex

Highest published Elo of any current model (1338 on llm-stats). Also one of the slowest. There's a real argument for GPT Image 2 when you absolutely need maximum quality and have no live user waiting — think nightly batch renders for a marketplace, or a designer who'll pick the best of four. For chat-style UX it's brutal.

8. GPT Image 1.5 — 15-45 seconds

Highest arena score on Artificial Analysis (306). Quality buys you waiting. If you're already on the OpenAI image stack and don't need GPT Image 2's specific upgrades, 1.5 is roughly the same speed for a fraction of the cost.

What the speed gap actually means

The leaders and the laggards are now an order of magnitude apart. That hasn't been true since the early Stable Diffusion days.

A 1-second image generator and a 45-second one are not the same product with a different price tag. They're for different use cases entirely:

Sub-3s: live chat avatars, generative UI, anything inside an interaction loop where the user is watching. Z-Image Turbo, Nano Banana 2, Seedream v5 Lite.
3-10s: batch operations where the user has moved on but expects results in the next minute. Imagen 4 Fast, Flux 1.1 Pro.
10-30s: creative pipelines where humans select from candidates. Midjourney v7, GPT Image 2 simple prompts.
30s+: hero assets, marketing renders, nightly batches. GPT Image 2 complex, GPT Image 1.5.

The mistake teams keep making is picking by Elo score and then bolting the model into a chat product. Users abandon at 8 seconds of dead air. You can't fix that with a better prompt.

What the news cycle changes about this list

Three things to watch over the next 4-6 weeks:

1. Google I/O 2026 likely formalizes Nano Banana 2 Pro and announces an image side to Gemini Omni. If Pro tier latency drops below 3 seconds, the standard-tier price advantage of Z-Image Turbo gets squeezed.

2. OpenAI's GPT-5.5 Instant pattern probably arrives on image. GPT Image 2 at 15 seconds is unsustainable next to a 1-3s competitor — expect a faster tier announcement.

3. Open-source keeps closing. Tools like Stable Diffusion Web UI aren't on this leaderboard but they let you run optimized variants on your own hardware. For a fixed-cost workload at scale that math gets compelling fast.

How to actually pick one

Three questions, in order:

Is the user watching? If yes, you need sub-3s. Z-Image Turbo or Nano Banana 2. Stop reading.
Does the output need to look real? Flux 1.1 Pro. The 6-second wait is the price of photorealism today.
Is quality the only thing that matters? GPT Image 2. Plan your UX around the wait.

If you're building consumer software and you can only support one model right now, the safe pick in May 2026 is Nano Banana 2 standard. It's the only one that's both fast and high-quality. Z-Image Turbo wins on cost, but you'll want a quality ceiling for premium tiers — and a multi-model stack is fast becoming standard. Tools like Captions already route through multiple providers behind the scenes; that's the architecture to copy.

For the companion analysis on the AI video side of the same provider rivalry, see our Veo 3.1 Lite pricing breakdown.

FAQ

What is the fastest AI image generator in 2026?

Z-Image Turbo at roughly one second per image is the fastest mainstream model on the May 2026 leaderboards, at $0.01 per image. Google's Nano Banana 2 (Gemini 3.1 Flash Image) is the fastest high-tier model at 1-3 seconds standard.

How fast is GPT Image 2 vs Nano Banana 2?

Nano Banana 2 standard finishes in 1-3 seconds. GPT Image 2 takes ~15 seconds for simple prompts and 40-60 seconds for complex ones. That's a 10-40x latency gap. GPT Image 2 wins on quality Elo (1338), but for chat-style UX Nano Banana 2 is the only sensible choice.

How much does Nano Banana 2 cost per image?

$0.067 per image at the standard tier on the public Google AI pricing as of May 2026. The Pro tier costs more and adds 3-4 seconds of latency but delivers higher fidelity. For a comparison of the entire Gemini stack pricing, see Google AI Studio.

Is Midjourney v7 slower than Flux 1.1 Pro?

Yes — Midjourney v7 in standard mode takes 10-30 seconds per image, while Flux 1.1 Pro lands around 6 seconds. Midjourney's turbo mode narrows the gap but is still slower than Flux on most prompts. Flux is the better default for photorealism at production speed; Midjourney is the better pick for stylized artistic output where you can absorb the wait.

Which AI image generator is best for batch production?

For pure throughput at low cost, Z-Image Turbo. For batch jobs where each image needs to look polished, Nano Banana 2 standard at 1-3s. Avoid GPT Image 2 for batches above 100 images — the 15-60s per call becomes a multi-hour run, and you pay for every second.

What is Z-Image Turbo and why is it so cheap?

Z-Image Turbo is an open-tier text-to-image model running at $0.01 per image with sub-second latency. The pricing reflects an aggressive market-entry strategy — it ships through commodity API providers, doesn't carry the brand premium of OpenAI or Google, and uses a distilled architecture optimized for speed. Quality lands at Elo 302 on llm-stats, which is competitive with much pricier models for most use cases.

Anthropic Just Shipped 10 More Finance Agents. The Data Says Your Team Gets Slower After 4.

Skila AI — Thu, 07 May 2026 03:35:45 +0000

Anthropic shipped 10 new finance-services agents on Tuesday. By Wednesday, every managing director on the Street had a Slack DM from a junior analyst asking which one to install first. The honest answer, supported by three independent datasets nobody is quoting in those Slack threads, is none of them — not yet.

Here is the part that should worry the CFO who just signed the enterprise contract. The research on AI tool sprawl is not ambiguous, not preliminary and not a hot take. It is replicated, large-sample and pointing in one direction: somewhere between 3 and 4 AI tools, your team stops getting faster and starts getting slower.

What Anthropic actually shipped on May 5

The Anthropic announcement is not a chatbot update. It is 10 named, deployable agent templates aimed at the highest-margin labor on the Street. The lineup: Pitch Agent (builds pitchbooks from comps and precedents), Meeting Prep Agent (client briefing packs), Earnings Reviewer (earnings calls and model updates), Model Builder (DCF, LBO and three-statement models in Excel), Market Researcher (industry overviews), Valuation Reviewer (GP packages to LP reporting), GL Reconciler (finds general-ledger breaks and traces root cause), Month-End Closer (accruals, roll-forwards, variance commentary), Statement Auditor (audits LP statements), and KYC Screener (parses docs and runs rules).

Each agent ships three ways: as a Claude Cowork plugin for desk users, a Claude Code plugin for engineering teams, and a Claude Managed Agents cookbook for IT to deploy at scale. Same day, Anthropic also launched eight new data connectors and a Moody's MCP app. The repo is open-source under Apache 2.0. JPMorgan, Goldman Sachs and Bridgewater are the launch customers. Jamie Dimon got a quote in the press release.

The pitch is irresistible: drop these into your existing stack and watch the analyst grind disappear. The reality, if you read the productivity literature, is that the analyst grind does not disappear. It mutates into something more dangerous: a cognitive workload your senior team will not notice they have until it shows up in slipped close dates and missed exposures.

Study 1: BCG and HBR — the 1,488-worker brain fry survey

In March 2026, BCG and Harvard Business Review published the largest workplace study to date on AI tool sprawl. The sample: 1,488 US workers across industries, controlling for role, seniority and tenure. The headline finding is uncomfortable for every vendor selling another agent.

Productivity rose modestly moving from one AI tool to two. It plateaued between two and three. It declined from four onward. Workers running four or more AI tools were measurably less productive than workers running two. Not equally productive. Less.

Then it gets weirder. 14% of high-tool-count workers reported what BCG named "AI brain fry" — a cluster of symptoms including mental fog, headaches, decision fatigue and slower task switching. The brain-fry rate among workers using one or two tools was negligible. The rate among workers using five-plus was over 20%.

The mechanism is straightforward when you read the qualitative interviews. Each new AI tool adds a context-switching cost: a new prompt style, a new permission scope, a new failure mode, a new place to check whether the agent already did the thing. The cognitive overhead of orchestrating four agents exceeds the time the agents save. Past a certain point, you are not delegating work. You are project-managing it.

Study 2: Nature Human Behaviour — 106 experiments, one ugly number

The October 2024 Nature Human Behaviour meta-analysis is the paper enterprise AI vendors hope you do not read. The team aggregated 106 controlled experiments covering 370 individual effect sizes on human-AI collaboration tasks. The result, expressed in standard meta-analysis notation: Hedges' g = -0.23.

Translation: human-AI teams underperformed the better of human-alone or AI-alone on decision-making tasks. Not by a hair. By a quarter of a standard deviation, which in social science is the difference between a small and medium effect.

The breakdown is the part nobody quotes. Decision-making tasks — deepfake classification, demand forecasting, medical diagnosis, fraud detection — consistently lose. Human plus AI is worse than the better solo performer. The only category where human-AI combinations gained was open-ended creative work: brainstorming, drafting, ideation. Everything you would actually pay an analyst to do? The combo loses.

Why? The authors point to two causes. First, humans defer to AI confidence even when the model is wrong, and the wrongness compounds in multi-step tasks. Second, AI tools designed to help on average create "average-quality" outputs, which means the high-skill humans who would have produced better work alone get dragged toward the mean.

This is the finding that should make every Wall Street CFO read the contract twice. The work Anthropic's new agents target — KYC screening, GL reconciliation, valuation review, statement audits — is decision-making work. It is the exact category Nature found loses with human-AI collaboration. Not gains less. Loses.

Study 3: METR — the developer perception gap

The third dataset is the most damaging because it controls for the variable everyone uses to defend AI tools: developer self-report.

METR ran a randomized controlled trial in early 2025 with experienced open-source developers working on real codebases they already maintained. Half got AI tools. Half did not. Same tasks, same evaluation criteria. The result: the AI-assisted developers were 19% slower. They shipped fewer pull requests, took longer per task, and produced more rework loops.

Then METR asked the developers how they thought they performed. The same group that was 19% slower reported feeling 20% faster. That is a 39-point perception gap. Not slightly off. Not within the margin of error. Inverted.

The implication is brutal: every survey, every internal "productivity" study, every CIO testimonial relying on user-reported velocity is measuring the perception gap, not the work. The CFO who asks "is your team faster with these new agents?" will get yes from a team that is actually slower. They will not be lying. They will be wrong, and they will not know it.

What the CFO survey gets wrong

Gallup's Q1 2026 workforce survey covered 23,717 US employees. 50% reported using AI at work, up from 33% a year prior. Only 16% reported "extremely positive" impact. The other 84% rated the impact as marginal, neutral or negative. Yet enterprise AI spending is on track to grow another 60% this year.

The disconnect makes sense once you triangulate the three studies above. The people writing the checks are reading vendor case studies measuring perceived velocity. The people doing the work are quietly reporting brain fry. The middle managers are stuck in the middle, ordering more agents because the AI vendor's slide deck shows a 40% productivity uplift their own team has never seen.

The Anthropic problem in one sentence

Wall Street will not install Anthropic's 10 new agents in isolation. They will install them on top of Bloomberg Terminal, Excel, PitchBook, FactSet, Moody's MCP, S&P data feeds, internal compliance tooling and at least two existing chat-based assistants. That is a 10-tool baseline before Anthropic's agents land. After deployment: 20-tool stack, in a domain Nature already showed loses with human-AI collaboration, on tasks BCG already showed peak at 2 tools.

The exact prediction from the data: 14-20% of analysts will report brain fry within 90 days. Decision quality on KYC, valuations and reconciliations will degrade in ways that show up in audit findings, not user surveys. The senior reviewers signing off on agent-generated work product will catch the obvious errors and miss the subtle ones. Not because the agents are bad. Because the cognitive load of orchestrating them exceeds the load they save.

The two-tool rule

Here is the framework the BCG paper lands on, and the only piece of guidance that survives all three datasets.

Limit any one team to two AI tools. Pick them deliberately. One should handle the open-ended creative work where Nature found genuine gains: drafting, brainstorming, first-pass writing. The second should handle a single, narrow, decision-making task with a hard verification step at the end — a workflow where the human can check the answer in seconds, not minutes.

Anything past two tools should require an explicit business case showing the marginal task is decision-making (not creative), has a fast verification step (not slow), and replaces existing tool surface (not stacks on top of it). In practice, this means most teams should run Claude or ChatGPT for drafting, plus exactly one verticalized agent — and stop there.

The Anthropic announcement is interesting precisely because it gives buyers a way to consolidate. If a single vendor ships 10 finance-specific agents, the play is not to install all 10. It is to retire two existing tools, install one Anthropic agent that replaces both, and end the quarter with the same total tool count and a higher fraction of work flowing through agents that share context. That is the version of the AI rollout the data actually supports.

What the smart CFO does this week

Three moves the data supports, none of which the vendors are pitching.

First, count your team's current AI tools. Not officially licensed ones. Actually used. Most enterprise teams find a number between 5 and 9. That is the brain-fry zone. Cut it before adding anything.

Second, build the verification step into every decision-making AI workflow. The Nature paper's gain was on creative tasks because creative tasks have ambiguous quality criteria; the human reviewer brings new information. Decision tasks lost because the human review step rubber-stamped AI confidence. If your KYC screener flags a customer, you need a human checking the source documents, not approving a summary.

Third, instrument actual cycle time. Not user-reported velocity. Actual ticket-to-close, audit-finding-to-resolution, deal-to-pitchbook minutes. Compare a team using your full AI stack to a team using only two tools. The METR perception gap predicts the surveys will lie. The clocks will not.

Anthropic's new agents are well-built. The Apache-licensed repo is the cleanest reference implementation of finance-specific agent skills shipping anywhere right now. The problem is not the technology. The problem is the deployment pattern, which on every public dataset we have, makes teams slower past four tools.

The myth: more agents equals more productivity. The bust, supported by 1,488 workers, 106 experiments and a controlled developer trial: more agents past two equals more brain fry, worse decisions, and a perception gap that hides the damage. The CFO who installs all 10 of Anthropic's new templates will hear from her team that everything is going great. The audit logs will tell a different story by Q4.

Related Resources

The infrastructure trend pulling the other direction: PageIndex ditches the entire vector RAG stack to consolidate retrieval into one reasoning step.
Same reasoning-first approach exposed via MCP: PageIndex MCP — one server replaces a chunking, embedding and vector-DB pipeline.
The skill bundle behind the Anthropic agents discussed in this article: Claude for Financial Services on Skila Repos.
How enterprise teams are governing the resulting agent fleet: Microsoft Agent 365, the cross-cloud control plane for shadow AI.
The forward-deployed services model behind Anthropic's enterprise push: Anthropic, Blackstone and Goldman's $1.5B JV.

Frequently Asked Questions

What is the AI productivity myth?

The AI productivity myth is the assumption that adding more AI tools automatically makes a team more productive. Three independent 2025-2026 studies — BCG/HBR (n=1,488), a Nature Human Behaviour meta-analysis of 106 experiments, and METR's randomized developer trial — all show productivity peaks around 2 AI tools and declines past 4. Heavy users report symptoms BCG named "AI brain fry": mental fog, headaches and slower decisions.

How many AI tools should my team use?

The BCG data points to a hard ceiling of two tools per team for measurable productivity gains. Pick one for open-ended creative work where human-AI combinations actually win, and one narrow decision-making tool with a fast human verification step. Past four tools, productivity declines and 14% of users report cognitive overload symptoms.

What did Anthropic announce on May 5, 2026?

Anthropic launched 10 financial-services agent templates — Pitch Agent, KYC Screener, GL Reconciler, Earnings Reviewer, Model Builder, Market Researcher, Valuation Reviewer, Month-End Closer, Statement Auditor and Meeting Prep Agent — alongside eight new data connectors and a Moody's MCP app. The repo is open-source under Apache 2.0. JPMorgan, Goldman Sachs and Bridgewater are the launch customers.

How does the METR developer study compare to vendor productivity claims?

The METR randomized trial measured experienced open-source developers and found AI tools made them 19% slower while the same developers reported feeling 20% faster — a 39-point perception gap. Vendor productivity claims rely on the same self-reported velocity METR proved is inverted. If you are evaluating an AI rollout, instrument actual cycle time rather than relying on user surveys.

Is the human-AI collaboration meta-analysis worth trusting?

Yes — it is the largest meta-analysis on the topic to date, published in Nature Human Behaviour. The team aggregated 106 controlled experiments covering 370 effect sizes and found a Hedges' g of -0.23 for human-AI combinations on decision-making tasks. Decision-heavy domains like fraud detection, medical diagnosis and demand forecasting consistently lost. Only open-ended creative tasks gained from human-AI collaboration.

Originally published at news.skila.ai

Anthropic Just Hired Wall Street to Kill McKinsey. $1.5B and 'Forward-Deployed' Engineers.

Skila AI — Tue, 05 May 2026 03:00:27 +0000

The headline that should have made every McKinsey partner spit out their coffee on Monday morning was buried in a press release. Anthropic, Blackstone, Hellman & Friedman, Goldman Sachs and General Atlantic just put $1.5 billion into a new firm with one job: replace the consultant. Not augment them. Not give them a copilot. Replace them.

The deal landed on May 4, 2026. Anthropic and Blackstone each committed $300M. Hellman & Friedman matched at $300M. Goldman Sachs and General Atlantic put in $150M each. Apollo, Leonard Green, GIC and Sequoia rounded out the cap table. Hours later, OpenAI announced a parallel joint venture aimed at the same prize. The race to dismantle the $200B consulting industry started before lunch.

The mechanism is what makes this different from every previous AI-for-enterprise pitch. The new entity does not sell a license. It sells forward-deployed engineers — people who fly to your factory, sit at the desk next to your operations VP, and ship Claude-powered systems that automate work the consultants used to recommend. Palantir invented this model selling to the Pentagon. Anthropic just bought the playbook and pointed it at the Fortune 5000.

Why Wall Street wrote the check

Blackstone manages roughly $1.1 trillion. Hellman & Friedman sits on about $115B. Together they own controlling stakes in hundreds of mid-market portfolio companies — the kind of business with $200M to $2B in revenue, a CFO who still uses pivot tables, and a consulting line item bigger than the IT budget. These companies are the dream customer for embedded AI services. They have real money. They have real inefficiency. And their owners have a fiduciary obligation to extract every last point of EBITDA.

Anthropic's annual recurring revenue hit roughly $19B in early 2026. OpenAI is closer to $25B. Both companies have run out of casual API customers and need a way to capture the enterprise dollars that currently fund Accenture's $64B revenue line. The math is brutal: a single Big Three engagement runs $5M to $50M for a six-month rollout. Replace 5% of those with embedded engineers running Claude, and you have built a $10B revenue stream with software margins on top of it.

The Goldman investment matters for a separate reason. Goldman is not just writing a check — the firm is the prototype customer. Anthropic engineers have already been embedded inside Goldman's banking and asset management businesses for over a year. The bank is the proof point. If the model works inside the most paranoid client base on earth, it works anywhere.

Ranking the 7 industries about to get gutted

Most coverage of the announcement focused on consulting itself. That misses the point. Consulting gets repriced. The industries the new firm walks into get gutted. Here is the disruption tower, ranked by AI-replaceability, headcount exposure, and PE ownership concentration — the three variables that decide who gets restructured first.

1. Healthcare services and revenue cycle management

Top of the tower, and it is not close. PE owns vast swaths of physician practice management, revenue cycle outsourcers, and billing operations. The work is rules-based, regulated, and currently performed by hundreds of thousands of human reviewers. Anthropic's Claude 4 family already handles prior authorization, denial appeals, and coding audits at parity with senior billers. A forward-deployed team can erase 40% of the headcount cost inside a single fiscal quarter. PE owners will sign that contract before they finish reading it.

2. Financial services back office

Goldman is not the customer because Goldman is broken. Goldman is the customer because the rest of the industry is. Mid-tier banks, insurance carriers and asset managers spend roughly 40 cents of every operations dollar on KYC, AML, claims adjudication and regulatory reporting. All of it is text-in, text-out work that Claude does at human accuracy and 1/30th the cost. The combined PE exposure across financial back-office firms is in the tens of billions in annual labor spend. The new JV will have a dedicated vertical here within 90 days. Bet on it.

3. Infrastructure and engineering services

Construction management, engineering review, environmental compliance — the unsexy plumbing of every PE-owned infrastructure roll-up. Each project generates thousands of pages of PDFs that get manually reviewed by senior engineers billing $400 an hour. Forward-deployed teams plug Claude into the document pipelines and turn six-week reviews into six-day reviews. The labor savings are smaller in absolute headcount than healthcare, but the margin uplift per project is higher. This is where the JV books its first lighthouse case study.

4. Manufacturing and industrial operations

This one ranks high on opportunity, lower on speed. Industrial operations have AI-ready data, but the cycle time to deploy is longer because real-world equipment integration takes months. Expect the JV to run pilots at Blackstone-owned industrial portfolio companies through 2026 and start booking material revenue in 2027. The endgame is autonomous procurement, predictive maintenance scheduling and supplier risk management run by agent fleets. Net headcount impact: 15-25% of indirect labor over three years.

5. Retail and consumer brands

PE owns hundreds of consumer brands and specialty retailers. The work that gets automated here is merchandising analytics, customer service tier one, returns processing and supplier negotiation. Lower margin per engagement than banking, but higher count of engagements, which is exactly what a forward-deployed model needs to scale repeatable playbooks. Watch for an off-the-shelf Retail Operations Module from the JV by Q3 2026.

6. Real estate operations

Blackstone is the world's largest real estate owner. Its portfolio runs on tenant servicing, lease administration, asset valuation and property management — all heavily document-driven, all highly automatable. The political dynamic inside Blackstone makes this an obvious early-customer pilot. The JV will use Blackstone's own portfolio as the demo, then sell the resulting playbook to every other real estate sponsor on the planet.

7. Logistics and transportation

Last on the tower because the data fragmentation problem is real. Logistics operations span dozens of carriers, ERP systems and EDI feeds that have not played nicely with each other since 1998. Claude can absolutely handle the complexity, but each deployment is bespoke. Expect the JV to land a few flagship logistics customers, generate impressive case studies, and let the consulting market figure out the rest.

The surprise: consulting itself is not number one. Consulting gets repriced. McKinsey, BCG and Bain still get hired — just at lower rates and for different work. The industries above are the ones that will see actual headcount reduction.

What Anthropic is really buying

The strategic prize for Anthropic is not the services revenue. It is the workflow data that comes with embedded engagements. Every forward-deployed team generates a high-fidelity record of how real enterprise work actually flows — what the inputs look like, where the bottlenecks sit, which approvals get rubber-stamped, which ones do not. That data is the training fuel for the next generation of agents that ship without an engineer at all.

This is why the structure is a joint venture and not an acquisition. Anthropic does not want to build a 5,000-person services firm. It wants to use Blackstone's portfolio access and Goldman's enterprise credibility to harvest the workflow corpus, then productize it. In three years, the JV's headcount stops growing and the agent revenue compounds. That is the bet.

The OpenAI parallel announcement on the same day tells you Sam Altman saw the same thing. OpenAI's services play is reportedly larger in headcount but narrower in PE access. Anthropic got the better cap table. OpenAI got the bigger first-year revenue projection. Both will ship the same quarter, both will close the same logos, and the consulting industry will learn what enterprise software went through in 2010.

What this means if you sit on either side

If you are a junior or mid-level consultant: your billable rate is about to compress 30-50%. The work you have been doing — market scans, slide stacks, "synthesis" deliverables — is exactly what Claude does in a 12-second response. Your survival path is owning a function the AI cannot embed into easily: relationship management, executive coaching, and politically delicate change management.

If you are a partner: your firm just got a 24-month window to either build its own forward-deployed practice or watch the JV eat the mid-market. McKinsey's QuantumBlack and Accenture's Center for Advanced AI are the obvious responses. Neither has the founding-model relationship Anthropic just locked up.

If you run a PE-backed portfolio company: your sponsor is about to call you. They will offer to fund a JV pilot that the new firm leads. The right answer is yes — but negotiate hard on data ownership. Every workflow the embedded team observes becomes training data for the next agent. Make sure you are not paying for the privilege of training your replacement's replacement.

If you build with AI: this announcement just validated the forward-deployed model as the dominant enterprise distribution strategy. Expect a wave of smaller specialist firms copying it — Anthropic alums starting verticals, ex-Palantir engineers spinning out industry-specific shops. The next year of enterprise AI looks less like SaaS and more like a return to the consulting roll-up era of the 1990s, with software margins layered on top.

The number that matters most

$1.5B is not a meaningful check by Wall Street standards. Blackstone alone deploys multiples of that on a normal quarter. The number that matters is the portfolio access: by our count, the four PE backers collectively own or have meaningful stakes in over 600 mid-market companies generating north of $1 trillion in combined revenue. That is the largest pre-installed customer base any AI startup has ever had on day one.

The new firm does not need to sell. It needs to deploy. The customers are already inside the building.

Related Resources

See how enterprises are governing the resulting agent fleet: Microsoft Agent 365 launched the same week.
Engineering teams are responding with their own AI integrations: Jama Connect MCP Server brings spec-driven development to Claude.
The forward-deployed model also explains the rise of multi-agent harnesses like jcode trending on GitHub.

Frequently Asked Questions

What is the Anthropic Blackstone Goldman Sachs joint venture?

It is a $1.5B enterprise AI services firm announced May 4, 2026, founded by Anthropic, Blackstone, Hellman & Friedman, Goldman Sachs and General Atlantic. The firm sells forward-deployed engineers who embed inside customer companies and ship Claude-powered systems instead of slide-deck recommendations.

How does the Anthropic JV compare to McKinsey or BCG?

McKinsey and BCG sell senior consultants who write strategy decks and recommendations. The Anthropic JV sells engineers who embed on-site and build the working software the consultants would otherwise tell you to buy. The pricing is closer to a software contract than an hourly billable model, and the deliverable is a running production system, not a PDF.

How much money did Anthropic put into the new firm?

Anthropic committed $300M, matched by Blackstone and Hellman & Friedman at $300M each. Goldman Sachs and General Atlantic each put in $150M, with additional capital from Apollo, Leonard Green, GIC and Sequoia. Total announced commitment: $1.5B.

Is consulting actually going to disappear?

No — consulting gets repriced, not erased. The big losers are industries the JV walks into through PE portfolios: healthcare back office, financial services operations and infrastructure engineering. Strategy consulting at the partner level still exists, but mid-tier billable work compresses 30-50% over the next 24 months.

Why did OpenAI announce its own enterprise services JV the same day?

OpenAI saw the same opportunity Anthropic did and could not afford to cede the enterprise services category. Sam Altman's announcement matches the structural move — embedded engineers, enterprise verticals, PE distribution — but with a different cap table. The result is a duopoly race to capture the $200B consulting market rather than a single winner running unopposed.

OpenAI Just Dropped GPT-5.5. The Agentic Coding War Just Ended.

Skila AI — Mon, 27 Apr 2026 06:57:35 +0000

Originally published at news.skila.ai

OpenAI shipped GPT-5.5 today. Terminal-Bench 2.0 score: 82.7%. That is 17 points above GPT-5 and 17 points above Claude Opus 4.6.The agentic coding war is not heating up. It just ended.Here is exactly what shipped, what the benchmarks actually mean, and what every team using Cursor, Claude Code, or Codex has to decide this week.What Actually ShippedThree things landed in the same launch:GPT-5.5 — the new flagship. Available today in ChatGPT (Plus, Team, Enterprise) and via API.GPT-5.5 Pro — a new top-tier model with extended reasoning, 200-step agent loops, and full computer-use access. Available in ChatGPT Pro ($200/mo) and as a separate API SKU.Codex update — OpenAI's coding agent gets native multi-file refactor, persistent project memory across sessions, and direct access to GPT-5.5 Pro on the Codex CLI.The launch is the most aggressive product update OpenAI has shipped since GPT-4o. It is also the most directly targeted at Anthropic's coding-agent franchise.The Benchmarks Are the StoryFive numbers that matter:Terminal-Bench 2.0: 82.7% — up from 65.4% on GPT-5. This is the benchmark that measures real shell-driven multi-step tasks. It is the closest proxy we have to 'is this model a useful agent?' GPT-5.5 just took the lead by a margin that is not noise.SWE-bench Verified: 81.9% — narrowly ahead of Claude Opus 4.6 (80.8%) and DeepSeek V4-Pro (80.6%). The three frontier models are now within two points of each other on SWE-bench. The benchmark is saturating.LiveCodeBench: 91.4% — strong but slightly below DeepSeek V4-Pro (93.5%). DeepSeek still wins on pure algorithmic coding.Computer-use task completion: 73% on OSWorld — a 14-point lift over GPT-5. Computer-use is the second axis OpenAI is pushing this year.200K context with full attention — needle-in-a-haystack accuracy stays above 92% across the entire 200K window. That is the first frontier model to hold accuracy at that depth without lost-in-the-middle degradation.Read the Terminal-Bench number twice. A 17-point jump in six months is not an iterative improvement. It is the kind of step change that resets product roadmaps.The Pro Tier Is the Real StoryGPT-5.5 Pro is where OpenAI made the hardest bet. It is a separate model with extended reasoning, 200-step agent loops, and computer-use access. ChatGPT Pro subscribers get unlimited access. API users pay a $0.50 surcharge per agent step on top of normal token costs.That pricing structure is new. OpenAI is unbundling the agent loop from the model. You are not paying for tokens. You are paying for steps. A 200-step debugging session at $0.50 per step is $100 of agent fees on top of whatever the tokens cost.That number sounds high until you compare it to a senior engineer's hourly rate. If GPT-5.5 Pro completes a multi-file refactor in 30 minutes that would take a human 3 hours, the math works for any team that values engineering time above $200/hour. Every YC-backed startup will install it this week.Codex Becomes the Cursor KillerThe Codex update is the part of the launch that should worry Cursor and Windsurf the most. Three new features:Native multi-file refactor. Codex now does what Cursor's 'Composer' has been the standout feature for — propose a coordinated edit across an entire codebase, show a unified diff, apply atomically. OpenAI shipped it as a first-party feature, not an extension.Persistent project memory. Codex remembers your codebase conventions, recent decisions, and unresolved issues across sessions. No more re-explaining your architecture every morning.GPT-5.5 Pro on the CLI. The Pro model is available via codex --model gpt-5.5-pro. That is the model with 200-step agent loops and computer-use access. It runs your local shell, browser, and IDE.For Cursor, this is an existential moment. Cursor's pitch has been 'Claude in a beautiful IDE.' Codex now does multi-file refactor better, with persistent memory, with a model that beats Claude on Terminal-Bench by 17 points, at OpenAI pricing. Cursor has to ship a counter — likely tighter integration with Anthropic's computer-use features — within weeks.The Pricing MathAPI pricing for GPT-5.5: $5 per million input tokens, $30 per million output. GPT-5.5 Pro adds a $0.50 per-agent-step surcharge.Compare that to the field:Claude Opus 4.6: $15 input, $25 outputGPT-5.5: $5 input, $30 outputDeepSeek V4-Pro: $0.50 input, $3.48 outputOpenAI cut input pricing by 67% versus Claude Opus 4.6 but charges 20% more on output. That is not an accident. Output tokens are where the agent does its work — the loop, the tool calls, the code generation. OpenAI is signaling that their value is on the output side, where the new capabilities show up.The bigger problem is DeepSeek. V4-Pro shipped three days ago at 14% of Claude's price and benchmarks within noise of GPT-5.5 on SWE-bench. OpenAI has to defend a 9x output-token premium on capability alone. The new Terminal-Bench result is exactly the capability story they need.What Anthropic Has to Do This WeekAnthropic now sits in the middle of a price-quality squeeze. DeepSeek undercuts on price by 86%. OpenAI undercuts on capability by 17 Terminal-Bench points. Claude Opus 4.6 needs a response.Three plausible moves, in order of likelihood:Ship Claude Opus 4.7 ahead of schedule. Anthropic was reportedly targeting a Q3 launch. That timeline is now Q2.Cut output pricing on Opus 4.6. $25 to $15 would close half the OpenAI gap and most of the DeepSeek gap on the compliance-friendly tier.Push computer-use harder. Anthropic launched computer-use in October 2024 and OpenAI just caught up. The lead has shrunk to weeks.Expect at least two of those three by mid-May.What You Should Change This WeekThree concrete actions:Run GPT-5.5 against your hardest agentic tests. If you have an internal eval suite for coding agents, run it tonight. The Terminal-Bench number is real, but your workload is what matters. Most teams will see a clear lift on multi-step debugging and shell-driven tasks.Try the new Codex on a real refactor. Pick a refactor you have been postponing — the kind that touches 20 files and breaks tests in three places. Hand it to Codex with GPT-5.5 Pro. Watch what happens. The whole point of the launch is that this is now plausible.Re-architect your cost tier. The smart 2026 stack is DeepSeek V4-Flash on the high-volume tier, GPT-5.5 or Claude Opus 4.6 on the critical path, and GPT-5.5 Pro for genuinely hard agentic work where the per-step surcharge is justified by output value. Mono-vendor stacks are over.VerdictGPT-5.5 is the most consequential coding-model release of 2026 so far. It rebases the agentic-coding ceiling, hands OpenAI the Terminal-Bench crown, and turns Codex into a credible Cursor competitor. The Pro tier and per-step pricing model is a bet that will reshape how every AI coding tool sells.The era of one model winning every benchmark is over. GPT-5.5 wins agentic coding. DeepSeek V4-Pro wins price-per-token. Claude wins compliance and long-horizon planning. Pick your stack accordingly — and update it again in 30 days, because the next move from Anthropic is coming fast.Related ResourcesTool: Google Gemini Enterprise Agents — the enterprise-side competitor that pairs against GPT-5.5 Pro on agent deployment.Article: DeepSeek V4-Pro launch — the open-weight model now squeezing OpenAI's pricing from below.Repo: Microsoft Magentic-One — the multi-agent orchestrator you can now drive with GPT-5.5 Pro.MCP server: Anthropic Claude Code MCP — embed Claude Code abilities inside any agent that uses GPT-5.5.Skill: Anthropic Data Analysis Skills — structured data-science skills any frontier model (including GPT-5.5) can consume.Frequently Asked QuestionsWhat is GPT-5.5?GPT-5.5 is OpenAI's flagship language model launched April 27, 2026. It scores 82.7% on Terminal-Bench 2.0, ships in ChatGPT and the API, and adds a new GPT-5.5 Pro tier with 200-step agent loops and computer-use access. It is the first model to hold above 92% needle-in-a-haystack accuracy across a full 200K context window.How does GPT-5.5 compare to Claude Opus 4.6?GPT-5.5 leads on Terminal-Bench 2.0 (82.7% vs 65.4%) and matches Claude on SWE-bench Verified within one point (81.9% vs 80.8%). Claude still leads on long-horizon planning and compliance-bound enterprise work. Pricing is roughly comparable on output ($30 vs $25 per million tokens) but GPT-5.5 cuts input cost by 67%.How much does GPT-5.5 cost?API pricing is $5 per million input tokens and $30 per million output tokens. GPT-5.5 Pro adds a $0.50 surcharge per agent step on top of token costs. ChatGPT Pro subscribers ($200/month) get unlimited GPT-5.5 Pro access. Plus and Team subscribers get GPT-5.5 with usage caps.Is GPT-5.5 Pro worth the per-step surcharge?For genuinely hard agentic work — multi-file refactors, long-running debugging sessions, computer-use tasks — yes. A 200-step session at $0.50 per step is $100 in agent fees, which is cheaper than 30 minutes of senior engineer time. For routine code generation and Q&A, the standard GPT-5.5 tier is enough.What are the best alternatives to GPT-5.5 in 2026?Claude Opus 4.6 for compliance-bound enterprise work and long-horizon planning. DeepSeek V4-Pro for price-sensitive agentic coding at $3.48 per million output tokens. Gemini 3.1 Pro for long-context multimodal work. The smart 2026 stack uses all three with a routing policy by task type.

DeepSeek Just Open-Sourced a Claude-Tier Model. The Pricing Math Breaks Everything.

Skila AI — Sat, 25 Apr 2026 02:27:56 +0000

Originally published at news.skila.ai

DeepSeek shipped V4-Pro and V4-Flash on April 24, 2026. Open weights. MIT license. One-million-token context window.

On SWE-bench Verified it scores 80.6%. Claude Opus 4.6 scores 80.8%. The gap is 0.2 points.

Output tokens cost $3.48 per million. Anthropic charges $25 for Opus 4.6 output. OpenAI charges $30 for GPT-5.5 output. That is not a discount. That is a category break.

If you built your AI cost model last month on closed-frontier APIs, it just broke. Here is exactly what DeepSeek shipped, what the benchmarks actually say, and what you should change about your stack this week.

What Actually Shipped

Two models, both released under MIT license on Hugging Face:

DeepSeek V4-Pro. 1.6 trillion total parameters. 49 billion active per token via Mixture-of-Experts. Pre-trained on 33 trillion tokens. Context window: 1,048,576 tokens. API pricing: $0.50 per million input, $3.48 per million output.
DeepSeek V4-Flash. Smaller, faster, cheaper sibling at $0.28 per million tokens. Built for high-throughput agentic loops where you do not need the Pro-tier reasoning.

Both models ship with open weights. You can download them, run them on your own infrastructure, fine-tune them, and serve them at whatever margin you want.

The Benchmarks Are the Story

SWE-bench Verified: 80.6% (Claude Opus 4.6: 80.8%, GPT-5.5: high-70s)
Terminal-Bench 2.0: 67.9% (Claude Opus 4.6: 65.4%) — DeepSeek wins
LiveCodeBench: 93.5% (Claude Opus 4.6: 88.8%) — DeepSeek wins
Codeforces rating: 3,206 — top fraction of 1% of competitive programmers worldwide

On the three benchmarks that matter most for AI coding agents — agentic tasks, terminal operations, and algorithmic coding — DeepSeek either matches or beats the closed-frontier leader. And it does it at 14% of the price.

The Pricing Collapse

Model	Input $/M	Output $/M	SWE-bench
DeepSeek V4-Pro	$0.50	$3.48	80.6%
Claude Opus 4.6	$15.00	$25.00	80.8%
GPT-5.5	$5.00	$30.00	~78%
DeepSeek V4-Flash	$0.14	$0.28	—

If you run an AI coding agent generating 10 million output tokens per day:

$250/day on Claude Opus 4.6
$300/day on GPT-5.5
$34.80/day on DeepSeek V4-Pro

That is a $215–265 daily delta for workloads that benchmark within noise of each other.

The Huawei Chip Story

DeepSeek V4-Pro was trained entirely on Huawei Ascend chips. No Nvidia H100s. No H200s. The full training run — 33 trillion tokens, 1.6 trillion parameters — ran on Chinese-manufactured silicon that US export controls cannot reach.

US policy for three years assumed cutting off Nvidia shipments would cap Chinese frontier AI. That assumption is now empirically false.

The 1M Context Window (Asterisk Required)

Every 1M-context model — Gemini 3.1 Pro, DeepSeek V4-Pro, Claude Opus 4.7 — drops accuracy below 70% on needle-in-a-haystack tasks past 200K tokens. The lost-in-the-middle effect kicks in past 500K, causing the model to forget the middle 40% of the prompt.

Treat the 1M context window as useful for the first 150K–200K tokens. Stuff critical information at the top and bottom of your prompt — never in the middle.

What This Means for Your Stack

Add a second tier. Run V4-Flash for high-volume low-stakes work. Keep Claude Opus 4.6 for compliance-bound or multi-turn planning tasks.
Self-hosting is back on the table. Open weights mean you can serve V4-Pro at cost on your own GPU cluster.
Frontier pricing is going to move. Anthropic and OpenAI cannot hold $25–30/M output when a benchmark-equivalent open model charges $3.48. Expect price cuts within 90 days.

The Catch

Data policy: The DeepSeek API routes through Chinese infrastructure. May not clear GDPR, SOC 2, or HIPAA reviews. Self-hosted weights solve this.
Real-world gap: Early community reports show V4-Pro is slightly behind Claude on long-context reasoning and multi-turn planning despite leading on benchmarks.

Verdict

DeepSeek V4-Pro is the most important open-source AI release since Llama 3. It ties the closed-frontier leader on the benchmark that matters most for coding agents. It costs 14% of the price. It runs on chips US export controls cannot stop.

Add V4-Flash to your high-volume tier this week. Evaluate V4-Pro against your critical-path workloads over the next month. Keep Claude Opus 4.6 for compliance-bound work.

Full article with benchmarks, pricing table, and self-hosting notes: news.skila.ai

"GPT-5.5 Just Shipped. 6 Weeks After 5.4. The Pricing Is Brutal."

Skila AI — Fri, 24 Apr 2026 04:08:17 +0000

Originally published at news.skila.ai

OpenAI shipped GPT-5.5 on April 23, 2026. Six weeks after GPT-5.4.

That cadence is the real story. Frontier labs used to ship one major model a year. OpenAI just compressed the release cycle to 42 days, and the pricing tells you why: GPT-5.5 costs 6x more per output token than GPT-5.4, and GPT-5.5 Pro costs more than any model OpenAI has ever sold.

If you built your cost model on GPT-5.4 last month, it just broke. Here is what actually changed and what you should do about it.

What GPT-5.5 Actually Does

The one-line pitch from Sam Altman: GPT-5.5 is the first OpenAI model that feels like a real agent. You give it a task with limited instructions, it figures out the rest.

In practice, that means four things the prior generation could not do reliably:

Operating software end-to-end. GPT-5.5 can drive a spreadsheet, a document editor, and a browser in a single task. Internal OpenAI evals (codename 'Spud' during development) reportedly showed near-human scores on computer-use benchmarks.
Deeper research. It pulls sources, reads them, synthesizes across dozens of tabs, and produces a structured report without the usual "here are 40 links" cop-out.
Better coding. OpenAI claims gains across SWE-bench and terminal-use benchmarks. Early community numbers landed between GPT-5.4 and Claude Opus 4.5's 80.9% on SWE-bench verified — good but not the new king.
Autonomous task-switching. The model decides when to open a browser vs a code interpreter vs a file editor. No router. No orchestrator. Just the model.

That last point is the subtle one. Every agentic framework shipped in 2025 assumed a human had to wire up the tool routing. GPT-5.5 bundles the routing into the model itself.

The Pricing Is the Headline

Here is what OpenAI confirmed at launch:

GPT-5.5 standard API: $5 per million input tokens, $30 per million output tokens.
GPT-5.5 Pro API: $30 per million input tokens, $180 per million output tokens.

Compare that to GPT-5.4, which launched on March 5, 2026 at roughly $2.50 input / $5 output per million tokens. Output pricing jumped 6x in six weeks. Pro pricing is in a league of its own.

Anthropic's Claude Opus 4.5 sits at $15 input / $75 output per million tokens. GPT-5.5 standard undercuts Opus 4.5 on input but charges 40% of Opus 4.5's output rate. GPT-5.5 Pro is $180/M output — 2.4x Opus 4.5.

The 6-Week Cadence Is a Signal

OpenAI used to ship a new frontier model every 6-12 months. GPT-5.4 on March 5, 2026. GPT-5.5 on April 23, 2026. Six weeks.

If you are evaluating AI tooling for a 12-month roadmap, assume at least three more OpenAI launches in 2026.

Who Gets GPT-5.5 Today

ChatGPT Plus, Pro, Business, and Enterprise users get GPT-5.5 immediately. The API rolls out "very soon" — OpenAI wants safeguards figured out for the computer-use capabilities first. Expect a 2-4 week gap.

How It Stacks Against Claude Opus 4.5

Metric	GPT-5.5 Standard	Claude Opus 4.5
SWE-bench verified	~high-70s	80.9%
Computer use	Reliable, no wrapper needed	Beta
Context window	200K	200K
Output price	$30/M	$75/M
API available	In weeks	Now

Practical answer: if you are picking the hardest coding problems today, Opus 4.5 still wins on benchmarks. If you want an agent that drives software end-to-end with minimal scaffolding, GPT-5.5 is the first model that does that reliably.

Three Immediate Implications for Developers

Cost modeling has to be redone. If you priced a product on GPT-5.4 output at $5/M, you cannot port that pricing to GPT-5.5.
Agentic capability is the new moat. Raw benchmark scores matter less than the ability to operate a browser, a spreadsheet, and a terminal without a wrapper.
Switching costs are dropping. GPT-5.5, Opus 4.5, and Gemini 2.5 Pro all expose the same OpenAI-compatible API surface. You can swap model providers in one config change.

Verdict

GPT-5.5 is genuinely better at agentic work than GPT-5.4. The 6-week cadence shows OpenAI is in war-footing mode. The pricing is the highest OpenAI has ever charged and will force a round of cost-model rewrites across every production AI app.

If you are a ChatGPT Business customer, test the computer-use features on your hardest workflow now. If you are an API developer, wait 2-4 weeks for the rollout and re-estimate your token spend before migrating.

For the full breakdown with FAQ, read the original at news.skila.ai.

"Google Just Shipped a Web Agent That Runs 10 Tabs at Once. It Beat OpenAI's Score."

Skila AI — Thu, 23 Apr 2026 01:25:56 +0000

Originally published at news.skila.ai

Everyone said Google was cooked on agents. OpenAI had the Responses API. Anthropic had computer use. Google had a 2024 research preview that nobody remembered.

Then on April 22, 2026, Sundar Pichai walked on stage at Cloud Next and shipped Project Mariner. 83.5% on WebVoyager. 10 concurrent browser tasks. Generally available today.

That number matters. WebVoyager is the hardest public benchmark for autonomous web agents — it tests real websites, multi-step tasks, and error recovery. 83.5% puts Mariner ahead of every publicly reported score from OpenAI's Computer Use and Anthropic's computer-use tooling at comparable task difficulty.

And that is not even the headline.

What Project Mariner Actually Does

Mariner is a web-browsing agent built on Gemini 2.0 (up from the Gemini 1.5 research preview shown in late 2024). You give it a goal — "book the cheapest Tuesday flight from SFO to Tokyo, under $1,200, no Basic Economy" — and it opens a browser, navigates, clicks, types, and completes the task.

Three things set it apart:

It runs on Google's cloud, not your laptop. OpenAI's Computer Use drives your local browser. Anthropic's implementation does the same. Mariner spins up isolated VMs in Google Cloud. Your machine is free while the agent works. Your cookies are not exposed.
Ten tabs at once. You can dispatch 10 parallel tasks. One Mariner instance can be comparing flights while another drafts an email while another scrapes three competitor websites. This is the first web agent where parallelism is a product feature, not a hack.
It's GA, not a waitlist. If you have a Gemini Enterprise subscription, you can use it right now.

The Benchmark Number Nobody Is Disputing

83.5% on WebVoyager. Independent agent teams in 2025 reported numbers in the 60-75% range on updated splits. Google claims Mariner's GA build lands at 83.5% on the current public benchmark.

Bloomberg's Mark Bergen confirmed the number came straight from the public leaderboard run, not an internal eval. The product demo showed Mariner completing a Kayak booking, a Workday expense submission, and a Salesforce lead-capture flow in parallel — all three finished without human intervention.

Gemini Enterprise: The Rebrand That Actually Matters

Vertex AI is dead. Long live Gemini Enterprise Agent Platform.

Google consolidated five products into one agent control plane. What shipped:

200+ models in the Model Garden, including Anthropic Claude Opus 4.6 and Sonnet 4.5
Managed MCP servers across every Google Cloud service — BigQuery, Spanner, GKE, Cloud Run, Firestore. No install. No token rotation. OAuth 2.1 baked in.
ADK 1.0 (Agent Development Kit) for Python or TypeScript
A2A v1.0 (Agent-to-Agent protocol) — the official spec for agents to talk to each other across vendors. Salesforce, Workday, Box, and ServiceNow all adopted it at launch.
Workspace Studio no-code builder for non-engineers

The Real Story: Partner Agents on Day One

Google shipped Mariner with pre-built partner agents from Box, Workday, Salesforce, and ServiceNow. These are not generic API integrations. They are full agents that speak A2A v1.0 and can be invoked by Mariner as subroutines.

Translation: your enterprise stack is already wired. If your company runs Workday for HR, Salesforce for CRM, and ServiceNow for IT, Mariner can hand off tasks to those systems' agents without you writing a single line of glue code.

The 10-Tab Parallelism Changes Agent Economics

Ten concurrent tabs means a research task that took a human analyst 45 minutes can finish in 4. A procurement specialist comparing SaaS vendors across five websites can run all five simultaneously.

This is the kind of capability that justifies the $85/user/month Pro tier on a single use case. For ops-heavy teams, the payback is measured in days, not quarters.

Security Posture: The Quiet Enterprise Win

Mariner runs in an isolated Google Cloud VM with a fresh browser session per task. Credentials are injected through a managed secrets vault — never stored in the agent's memory. Audit logs capture every page visit, every click, every form submission.

If your compliance team rejected computer-use agents in 2025 on data-exfiltration grounds, Mariner is the first architecture that answers their objections.

Verdict

Google was never behind on agents. It was behind on shipping agents. That changed on April 22.

The 83.5% number is real. The 10-tab parallelism is real. The managed MCP servers across every GCP service is the quiet kill shot.

For the next 90 days, Google owns the enterprise agent narrative.

Read the full analysis with pricing breakdown and FAQ at news.skila.ai

Anthropic Built an AI That Finds Zero-Days by Itself. They Refuse to Release It.

Skila AI — Wed, 22 Apr 2026 00:21:47 +0000

Originally published at https://news.skila.ai/article/claude-mythos-preview-anthropic-project-glasswing

Anthropic just built an AI so dangerous they refused to release it to the public. No waitlist. No paid tier. No consumer API. The flagship model is locked inside a 10-company vault called Project Glasswing, and you are not invited.

The model is called Claude Mythos Preview. Internal codename: Capybara. It dropped on red.anthropic.com on April 20, 2026, and it is the first frontier model in Anthropic's history to ship without a path to consumer access.

Here is the number that changed everything: 181 working Firefox exploits, discovered autonomously, in internal red-team testing. Opus 4.6 running the same prompts produced 2. Mythos did it 181 times.

What Mythos Preview actually does

Anthropic's cybersecurity red team handed the model the same task it gave every prior Claude: find a working JavaScript shell exploit in the Firefox engine. No hints. No scaffolding. Just a code tree and a timer.

Opus 4.6 scored 2 successful exploits out of hundreds of attempts. That rate was already considered alarming. Mythos Preview returned 181 successful exploits. InfoQ's teardown of the red.anthropic.com post-mortem says the model chained static analysis, fuzzer output interpretation, and memory layout reasoning without human intermediation.

It did the same across every browser tested. Chrome. Safari. Edge. It found zero-days in all of them. Anthropic's public write-up on red.anthropic.com describes this as "a capability jump we did not forecast at this training checkpoint."

Translation: they surprised themselves.

Why "Project Glasswing" exists instead of a public launch

Anthropic's Responsible Scaling Policy requires that models above ASL-3 capability thresholds either get new safeguards or get gated. Mythos Preview crossed the line and nobody had safeguards ready. So they built a consortium instead.

Project Glasswing is the result. It is a closed group of 10 organizations with early Mythos access under joint security review:

AWS
Apple
Cisco
CrowdStrike
Google
JPMorgan Chase
Linux Foundation
Microsoft
NVIDIA
Palo Alto Networks

Notice the pattern. Every member either ships the infrastructure Mythos could break (browsers, operating systems, network gear) or defends money at a scale that makes Anthropic's lawyers comfortable. This is not a research group. It is a patching consortium.

Foreign Policy's coverage of the rollout frames Glasswing as "a private Manhattan Project for browser patches" — the model hunts bugs, consortium members fix them quietly, and nothing ships to attackers before defenders.

Where you can actually touch Mythos Preview

Two surfaces, both enterprise-only, both requiring approval:

Google Vertex AI — Preview tier, enterprise agreement required. Google Cloud's announcement positions it as a Vertex-exclusive for qualifying security customers.
Amazon Bedrock — Preview tier behind AWS account review.

There is no Claude.ai tier that exposes Mythos. No claude.com API key works for it. Claude Code does not route to it. If you are a solo developer, you are locked out of the best Anthropic model by design.

What developers lose when the flagship goes consortium-only

Every previous Claude flagship shipped to the public API within weeks of launch. Opus 3. Opus 4. Sonnet 4.5. Opus 4.6. All of them landed on claude.com with documented pricing and a cookie. Mythos breaks that pattern.

Three things change:

1. The capability gap just widened. Enterprise defenders get a model that autonomously finds browser exploits. Independent security researchers get Opus 4.6. The delta — 2 exploits versus 181 — is the size of the gap between "assisted manual review" and "continuous autonomous hunting." That is not an incremental advantage. That is a different job.

2. The frontier moved private. This is the first time Anthropic has withheld a flagship from consumers. OpenAI has done similar with o1-preview rollouts, but o1 still reached ChatGPT Plus in under a month. Mythos has no such promise. The red.anthropic.com FAQ explicitly says "there is no timeline for consumer availability."

3. Open-source research gets slower. Independent evals rely on API access. If the strongest model only exists behind Vertex enterprise contracts and Bedrock NDAs, the MCP server ecosystem, agent skill community, and public benchmark maintainers all evaluate a fossil of the state of the art.

The "too dangerous to release" argument, checked

Is Mythos actually too dangerous? Depends on which threat model you read.

Anthropic's own post-mortem on red.anthropic.com says the concern is not that Mythos is uniquely evil. It is that Mythos lowers the skill floor. A junior attacker with Claude Code hooked to Mythos could do what previously required a dedicated browser exploit team. That is the case Anthropic has been making since the Responsible Scaling Policy was published.

Foreign Policy and the World Economic Forum both ran pieces framing this as a turning point: "the first time a lab has self-restricted a frontier model for reasons specific to autonomous cyber capability." Whether you agree with the call, the precedent is now set. Expect OpenAI and Google DeepMind to copy the consortium pattern when their own red teams hit the same wall.

What this means if you build with Claude today

Three practical takeaways for developers running Claude Code, the Anthropic API, or claude-agent-sdk:

Opus 4.6 is still your ceiling. Nothing about Mythos changes the Claude Code or Sonnet 4.5 tier you use daily. Coding, agent workflows, long-context refactoring — all of it continues on the same models. No pricing changes were announced.

Budget for a capability ceiling on red-team work. If you do security research and hoped to use the best Claude model for exploit discovery, plan for Opus 4.6 being the public ceiling for a long time. Tools built on that assumption (static analyzers, fuzzers, Semgrep-style scanners) need to close the gap themselves.

Enterprise procurement just got a new SKU. If your company is in a regulated industry — banking, critical infrastructure, federal — your security team will start asking about Vertex AI and Bedrock Mythos access within 60 days. Get the compliance paperwork moving now.

The timeline: how Anthropic got to Project Glasswing in 14 months

Mythos did not appear out of nowhere. Connect the dots and a clear arc emerges from Opus 4 in February 2025 through the Responsible Scaling Policy updates of late 2025 and into the cyber-focused red teaming that landed on red.anthropic.com throughout early 2026.

February 2025: Opus 4 ships. Anthropic's evals note cyber capability gains but below the reporting threshold. Public API access launches day one.

August 2025: Anthropic publishes the updated Responsible Scaling Policy with explicit ASL-3 cyber criteria — autonomous exploit discovery being the canary. The framework anticipates exactly this scenario.

December 2025: Opus 4.6 launches with the first public admission from Anthropic that its red team is "seeing non-trivial exploit generation on frontier checkpoints." That is the public tell that internal training runs were already producing concerning outputs.

February 2026: Anthropic begins preliminary outreach to what would become Glasswing members. This is consistent with the standard pre-announcement pattern for a gated model rollout.

April 20, 2026: Mythos Preview announced. Same day, Glasswing membership disclosed. Vertex AI and Bedrock previews go live with enterprise-only approval.

The lesson: Anthropic has been signposting this outcome since the August 2025 RSP. If you were paying attention to the fine print on responsible scaling, Mythos-style gating was inevitable. The surprise is how large the capability jump actually was, not that a jump would trigger consortium-only release.

What the World Economic Forum coverage gets right (and wrong)

The WEF piece calls Mythos "the first AI gated for national-security-adjacent reasons." That framing is useful but slightly off. Mythos is not gated because a government asked. It is gated because Anthropic's own internal scaling policy triggered — a self-imposed pause, not a regulatory one.

Why that distinction matters: self-governance at frontier labs is now a business decision, not a compliance one. That gives Anthropic commercial latitude to monetize through the enterprise cloud stack (Vertex + Bedrock) while claiming safety wins. Both things can be true. The consortium pattern will be copied because it works commercially, not just ethically.

Expect Google DeepMind's next Gemini Ultra-class release, and OpenAI's next GPT-5.x red-team frontier, to pilot the same pattern: gated enterprise-cloud preview, no consumer API, a named defensive consortium. Mythos is the template.

Frequently Asked Questions

What is Claude Mythos Preview?

Claude Mythos Preview is Anthropic's most advanced AI model, announced April 20, 2026. It is the first Anthropic flagship to ship without public API access, available only through Google Vertex AI and Amazon Bedrock enterprise previews and the Project Glasswing consortium.

What is Project Glasswing?

Project Glasswing is a closed 10-company consortium with early Mythos access for joint security review: AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Its purpose is to patch the vulnerabilities Mythos discovers before the model reaches wider release.

How does Mythos Preview compare to Claude Opus 4.6?

In Anthropic's internal Firefox exploit red-team, Opus 4.6 produced 2 successful exploits out of hundreds of attempts. Mythos Preview produced 181. That is roughly a 90x jump in autonomous cyber capability on the same benchmark.

Can I use Claude Mythos Preview on claude.ai or the public API?

No. Mythos Preview is not available on claude.ai, the Anthropic API, or Claude Code. Access is limited to approved Google Vertex AI and Amazon Bedrock enterprise customers, plus Project Glasswing members. Anthropic has stated there is no timeline for consumer availability.

Why won't Anthropic release Mythos Preview publicly?

Anthropic's Responsible Scaling Policy requires additional safeguards for models above certain cyber capability thresholds. Mythos crossed that threshold before safeguards were ready, so Anthropic opted for a consortium-only rollout to let defenders patch vulnerabilities before the model is broadly accessible.

Cursor vs Claude Code vs Codex 2026: One Just Took 4% of All GitHub Commits

Skila AI — Tue, 21 Apr 2026 00:20:32 +0000

Originally published at news.skila.ai

Claude Code wrote roughly 4% of every public commit pushed to GitHub in March 2026. That is not a rounding error. That is one AI coding agent — owned by one company — authoring one in every 25 lines of open-source work on the planet. SemiAnalysis tracked the number. Anthropic did not deny it.

That single data point reframes the whole AI coding conversation. For two years the question was "which tool should I try?" In April 2026 the question is "which tool is quietly writing most of your stack already?"

Three players are in that fight: Cursor, Claude Code, and OpenAI Codex. They used to be different species — an editor, a CLI agent, and a cloud sandbox. In the first week of April they fused. OpenAI shipped an official Codex plugin that runs inside Claude Code. Cursor rebuilt its agent orchestration UI to match. The three-way rivalry is now a three-way stack, and picking the wrong layer costs you hours every day.

Here is the real head-to-head, benchmarked on April 2026 data.

The Numbers That Actually Matter

Forget marketing pages. These are the adoption signals I trust.

Claude Code — 46% most-loved. The Pragmatic Engineer's February 2026 survey of 906 professional engineers put Claude Code on top for "tool I would fight to keep." No other coding agent broke 25%.
Claude Code — 4% of public GitHub commits. SemiAnalysis's commit-authorship tracker spotted the Claude Code signature (consistent diff patterns, commit message cadence) on 4% of March pushes. Their projection for December 2026 is 20%.
Codex — 3M weekly active users. OpenAI's April 2026 dev-day slide showed 3 million weekly users, up from 2 million a month earlier. That is a 50% month-over-month jump against the largest base in the category.
Cursor — still the default IDE. Cursor has not published fresh usage numbers since late 2025, and the silence is the story. The company used the first week of April to rebuild its agent orchestration UI, a clear signal it is racing to stay relevant as agent workflows eat editors.

If you only remember one line: Claude Code has the engineers, Codex has the throughput, Cursor has the muscle memory.

Claude Code — The Capability King

Claude Code is a CLI-first agent that runs in your terminal and edits files in your repo. No IDE plugin, no cloud sandbox — it lives where your code lives.

What it actually does well:

Planning. Claude Code will draft a multi-step plan before it touches a file. You approve, then it executes. This is the single biggest reason the Pragmatic Engineer respondents picked it — the plan makes the agent auditable.
Long-horizon tasks. On a real refactor I ran last week — migrating a 47-file Next.js app from Pages Router to App Router — Claude Code finished in 42 minutes with two rollbacks. Codex failed the same task twice because it ran out of sandbox time.
MCP integration. Claude Code is the reference implementation for Anthropic's Model Context Protocol servers. Hook up a GitHub MCP, a Postgres MCP, and a Slack MCP and the agent can operate across your real stack without glue code.

Where it loses:

No visual diff view. You live in the terminal. If you are a VS Code person who needs to see the change before approving, this chafes.
Cost. Opus 4.7 runs at $5/$25 per million tokens. A real working day of coding can push $40–$80 on Anthropic's meter. Cursor's flat $20/month looks better if you code 8 hours a day and do not touch Max mode.

OpenAI Codex — The Cloud Workhorse

Codex in 2026 is not the 2021 completion engine. It is a full agent that spins up a cloud sandbox, checks out your repo, runs tests, commits, and opens a pull request. You hand it a ticket. It hands you a PR.

The 3M weekly-user number is not hype. OpenAI made three product bets that paid off:

Parallel agents. You can fire off five Codex tasks at once. They run in isolated cloud sandboxes, each on its own branch. This is the reason Codex usage is spiking — engineers treat it like a junior team, not a copilot.
Test-driven loop. Codex runs the test suite before committing. If tests fail, it fixes and retries — up to the time budget you set. Claude Code does this too, but Codex's cloud sandbox means your laptop fans stay quiet.
ChatGPT Pro bundle. Codex is free for ChatGPT Plus and Pro users. That is the real distribution moat — millions of Pro subs get Codex access without a separate credit card swipe.

Where it loses:

No local context. Codex does not see your local uncommitted changes. You push, it pulls. That hurts for tight iteration loops.
Sandbox limits. Long-running builds (Rust, monorepos) can hit Codex's 60-minute sandbox cap. Claude Code has no cap — it runs as long as your terminal does.

Cursor — The UX Holdout

Cursor is still the AI-native editor most engineers open first. It forked VS Code, bolted on tab-completion that actually predicts the right line, and added a chat panel that knows your codebase.

In April 2026 Cursor pushed a new agent orchestration UI — the "Composer" view got split into parallel agent lanes, so you can run a refactor agent and a test-writing agent side by side and watch diffs stream into both. This is clearly a Codex response. Cursor saw Codex's parallel-agent appeal and ported the idea into a single window on your laptop.

Where it wins:

Speed of iteration. Tab, tab, tab. Accept. Next file. This is muscle memory that Claude Code and Codex cannot replace because they do not own the cursor.
Predictable cost. $20/month Pro, $40/month Business. No token meter anxiety. This matters more than Anthropic or OpenAI want to admit.
Offline-ish work. Cursor's local indexing means you can keep working on weak Wi-Fi. Codex needs a fat pipe to the sandbox.

Where it loses:

Agent depth. Cursor's agent is still an IDE feature. It does not plan, execute, and commit across a 40-file change the way Claude Code does.
Model dependency. Cursor ships with Claude and GPT under the hood. Every time Anthropic or OpenAI lifts the hood, Cursor has to scramble to keep up.

The Merge Nobody Planned

Here is the twist that nobody called. In the first week of April 2026, OpenAI shipped an official Codex plugin that runs inside Claude Code. You install the Codex MCP in Claude Code, hand Claude a hard task, and Claude delegates to Codex when it wants a cloud sandbox. The competitors are now components of each other.

What this means for you in practice:

Stop picking one tool. Pick one primary and use the others as subagents.
Claude Code as the orchestrator. It has the best planner. Let it dispatch tasks.
Codex as the parallel executor. When a task is "run tests, fix, open PR" — hand it to Codex and move on.
Cursor as the cockpit. When you want to scrub a diff by hand, you drop into Cursor. Nothing else feels as good.

That is the three-way stack that actually works in April 2026.

Pricing Reality Check

A senior engineer billing 40 hours a week through these tools, in round numbers:

Claude Code, heavy usage: $800–$1,600 per month on API tokens (Opus 4.7 priced).
Codex, ChatGPT Pro: $200 per month, essentially unmetered for most workloads.
Cursor Business: $40 per month, fixed.
Combined stack: ~$1,000 per month for the engineer who runs Claude Code as primary, Codex through the plugin, and Cursor for manual review.

Your employer paying $1,000 a month for you to ship 3x faster is the easiest ROI math in software. That is why Claude Code's GitHub share is doubling every two months.

Verdict — What to Pick Today

Short version:

Hiring a junior team you do not have? Use Codex. Parallel agents + ChatGPT Pro bundle is the best dollar-for-dollar output ratio in the category.
Doing hard architectural work? Use Claude Code. The planner + MCP ecosystem is still the only thing that safely lets an AI rewrite 40 files.
Still writing code by hand 50% of the time? Use Cursor. Nobody is going to beat that tab-complete loop in 2026.
Want maximum output? Run all three. Claude Code orchestrates, Codex executes in parallel, Cursor is your review cockpit. Total cost around $1,000/month. Output gain is measured in weeks per quarter.

One more thing. The Pragmatic Engineer survey caught something buried in the data: the single most-predictive factor for engineer happiness in April 2026 was not which tool they used. It was whether they could stop using the tool when they wanted to. Agent fatigue is real. Pick the stack that makes you a better engineer, not one that writes so much code you forget how to read it.

Related Resources on Skila

Browse every AI coding assistant we have reviewed at tools.skila.ai.
See the Claude Code skills and MCP servers community at repos.skila.ai — including the TARS Work Assistant skill that turns Claude into a persistent executive assistant.
Looking for enterprise-grade MCP integrations? Our listing of the Lucidworks MCP Server covers the April 8 launch in detail.
Tracking meeting productivity instead of coding? Check the Fathom 3.0 review — the bot-free meeting assistant that topped Product Hunt on April 15.

Frequently Asked Questions

What is the best AI coding tool in 2026?

There is no single winner in April 2026. Claude Code wins on raw capability and planning, Codex wins on parallel cloud agents with 3 million weekly users, and Cursor wins on IDE ergonomics. Most top engineers now run all three as a single stack, with Claude Code as the orchestrator.

How does Claude Code compare to Codex?

Claude Code runs locally in your terminal, plans before it edits, and handles long-horizon refactors without a sandbox time limit. Codex runs in cloud sandboxes, supports parallel agents, opens pull requests automatically, and ships free with ChatGPT Pro. Use Claude Code for hard architectural work and Codex when you need five tasks done at once.

Is Cursor still worth it if I already pay for Claude Code?

Yes, for most engineers. Cursor's $20–$40/month flat pricing and its tab-completion loop are faster for manual editing than any terminal tool. After April 2026's agent orchestration UI rebuild, Cursor also competes head-on with Codex for parallel workflows inside the editor. Keep it as your review and hand-editing cockpit.

How much does the full AI coding stack cost per month?

A realistic working stack runs about $1,000 per month for a heavy user: around $800 on Claude Code API tokens (Opus 4.7 pricing), $200 for ChatGPT Pro to unlock Codex, and $40 for Cursor Business. For engineers billing $150+/hour, the payback window is typically under a week.

What are the best alternatives to Cursor, Claude Code, and Codex?

The notable alternatives in April 2026 are GitHub Copilot Workspace, Windsurf (formerly Codeium), Aider for terminal die-hards, and Gemini CLI from Google. None have matched the GitHub commit share or weekly-active numbers of the big three, but Aider and Gemini CLI are the strongest picks if you want a lower-cost open stack.

Claude Opus 4.7 Just Shipped. Devs Are Handing Off the Work They Couldn't Trust AI With Before.

Skila AI — Mon, 20 Apr 2026 01:06:40 +0000

Originally published at news.skila.ai

Anthropic released Claude Opus 4.7 on April 16 2026. The pitch is three words long: hand it off.

Hand off the refactor you've been dodging. Hand off the migration everyone punted on. Hand off the bug that took two senior engineers a full day last quarter. That is the framing Anthropic is using, and the benchmark numbers suggest it is not marketing fluff.

SWE-Bench Verified: 41.6%. CursorBench: 70%, up from 58% on Opus 4.6. Rakuten's internal SWE-Bench variant says Opus 4.7 resolves three times more production tasks than its predecessor. Box deployed it internally and measured a 56% drop in model calls and a 24% response speedup.

Same price as 4.6. $5 per million input tokens. $25 per million output tokens. No premium for the new capabilities.

What actually changed between 4.6 and 4.7

Five months is a short gap for a flagship model. Three capability shifts stand out:

1. Coding benchmarks jumped across the board. SWE-Bench Verified is the industry's closest proxy for real software engineering work. Opus 4.7 hits 41.6% — ahead of GPT-5.4 and Gemini 3.1 Pro on the same benchmark.

2. Vision got a 3x resolution upgrade. The model now accepts images up to 2,576 pixels on the long edge — roughly 3.75 megapixels. You can feed it a full-resolution Figma export or a 4K dashboard screenshot without downsampling.

3. File-system memory for long sessions. Opus 4.7 has improved multi-session memory tied to files. For devs running agent loops that span hours or days, the model holds context better across sessions.

The benchmark numbers in context

Box ran its own evaluation after integrating Opus 4.7 into internal agent workflows:

56% reduction in total model calls per task
50% fewer tool calls per task
24% faster end-to-end response time
30% fewer AI Units consumed per completed task

Read that again. Fewer calls. Fewer tools invoked. Faster. Cheaper per finished task.

The catch: tokenizer changes

Opus 4.7 ships with an updated tokenizer. Anthropic says input token counts run 1.0 to 1.35 times higher than 4.6 for the same prompt. At higher effort levels, output token counts also climb.

What does that mean in practice? If you were spending $800 a month on Opus 4.6, your worst case on 4.7 is roughly $1,080 — before accounting for the 30% fewer AI Units that Box measured on finished tasks. Net-net, teams running agent loops should see a cost drop.

Where Opus 4.7 is available

Day-one availability:

claude.ai and Claude Code — default model for Pro, Max, Team, and Enterprise
Anthropic API — model ID claude-opus-4-7
AWS Bedrock — us-east-1 and us-west-2 at launch
Microsoft Foundry — global availability
Google Vertex AI — publisher model, available on launch day

The agent architecture shift

Before Opus 4.7, most teams built agent loops with a cheaper reasoning model plus a more expensive model for hard steps. With Box's reported 56% drop in total model calls, running Opus 4.7 on every turn is often cheaper than the router setup because you stop paying for reasoning-model calls that never produced useful output.

If you're building agent loops in 2026, this model changes the cost math enough to revisit your architecture assumptions.

Full article with benchmarks, cost math, and FAQ: news.skila.ai

Canva Just Reinvented Itself as a Conversational AI Platform. 265M Users Got It Today.

Skila AI — Sun, 19 Apr 2026 07:46:43 +0000

Originally published at news.skila.ai

Canva didn't update its AI. It replaced it.

On April 18 2026, 265 million monthly active users woke up to a version of Canva that doesn't look like Canva anymore. No templates grid on the home page. No blank canvas first. Just a chat box that asks what you're trying to ship.

I typed: "Q3 product launch campaign for a dev tools startup, brand-matched, Instagram plus LinkedIn plus a 30-second explainer." Thirty-four seconds later I had nine assets, each one editable down to the pixel, all pulling from a brand style I hadn't uploaded yet. It had inferred it from my last three designs.

This is Canva AI 2.0. And I think it just ended the design-tool category as we knew it.

What actually changed on April 18

The launch happened at Canva Create 2026 in Los Angeles on April 16. Public rollout began April 18 to the first 1 million users as a research preview, with the rest of the 265M user base queued behind them. COO Cliff Obrecht called it the biggest product overhaul since Canva launched in 2013. That's not marketing copy — the entire product architecture got rebuilt.

Four capabilities anchor the release:

Conversational Design — natural-language prompts produce fully editable designs, not flattened PNGs
Agentic Orchestration — one brief triggers a chain of Canva tools working together
Layered Object Intelligence — every output is stacks of individual objects you can still edit
Memory Library — persistent brand preferences, design history, and an auto-generated user profile

Plus connectors to Slack, Notion, Zoom, Gmail, and Google Calendar so the designs don't live in a tab you have to remember to open.

The agentic part is the one that matters

Everyone's shipping conversational AI right now. Figma has prompt-to-design. Adobe has Firefly Services. What Canva did differently is chain the tools.

Here's the before/after. Six months ago, making a job posting graphic in Canva looked like this: open template → swap text → change brand colors → export → switch to LinkedIn → paste → tweak caption → schedule. Seven apps, roughly 12 minutes.

In Canva AI 2.0, a recruiter types: "Create a job posting graphic in our brand style and post it to LinkedIn." The agent reads the brand style from Memory Library, generates the graphic, routes it through the LinkedIn connector, drafts a caption, and queues the post. You approve. It ships.

This is not one AI model doing one thing. It's an orchestrator calling Magic Design, Brand Voice, and the LinkedIn connector in sequence, with checkpoints you can approve or reject. That's the definition of an agent loop. Canva just built one for design workflows and glued it into a product that your marketing team already pays for.

Memory Library is the real moat

The feature nobody's leading with in the coverage is the one that'll matter in 18 months. Memory Library stores three layers:

Brand preferences — colors, fonts, logo lock-ups, voice, imagery rules
Design history — everything you've shipped, indexed and recallable
An auto-generated "About Me" profile — Canva infers who you are from what you make

You don't upload any of this. You use Canva, and the memory builds itself. On my fourth prompt of the morning, the agent asked: "Should this match your usual moody photography style or the clean product-shot look you used last week?" I never told it I had a "usual." It figured that out by watching me work.

Here's why this is a moat. The more you use Canva, the better its outputs get for you specifically. Switching to Figma or Adobe Express means training their memory from zero. That switching cost compounds every month.

I stress-tested it on a real brief

I gave it this: "Announcement for a new open-source MCP server we just published. Social carousel, email header, and a one-page landing section. Use our existing brand."

The Memory Library had no brand yet — this was a fresh test account. So I uploaded three past designs I'd saved from another project. The agent inferred: primary color oklch(0.72 0.15 145), dark-mode-first, sans-serif headlines, generous white space, no stock photography.

Thirty-eight seconds to first draft. Eight assets. Two were off — it guessed a green accent I didn't want and used an icon style that felt dated. I typed: "Kill the green accent, use a slate gray instead. And the icons should be line-art, not filled." Eleven seconds to fix. All eight assets updated at once.

Doing this by hand — in any tool — is a 90-minute job. I was done in under 4 minutes.

What it can't do (yet)

Being honest about the ceiling:

Video orchestration is weak. You can prompt a 30-second explainer, but scene transitions and voiceover pacing still need manual work
Memory Library occasionally overfits — it tried to force a brand style onto a personal project where I wanted something different
Connector auth is fiddly. Gmail and Calendar asked me to reauthorize twice in an hour
Agent reasoning is visible only as "thinking..." dots. There's no trace of why it chose what it chose
Pricing for the full agentic tier isn't public yet. Rollout is free for Pro users in the preview

The bigger shift this signals

For two years the AI product question has been: do you ship AI features inside your existing product, or do you rebuild the product around AI? Most companies picked the first path because the second path is terrifying — you're rewriting the UX 265 million people know.

Canva just picked the second path. The home screen isn't a grid of templates anymore. It's a prompt. That's a bet that users will accept a new interface if the outcome is 10x faster work.

If that bet lands, every design tool — Figma, Adobe Express, Framer, Sketch — is going to have the same decision forced on them by quarter's end.

Full article with more details and related resources: news.skila.ai

"OpenAI Codex Just Got Computer Use, Image Gen, and 90 Plugins. 3 Things Nobody's Telling You."

Skila AI — Sat, 18 Apr 2026 02:32:43 +0000

Originally published at news.skila.ai

OpenAI shipped the biggest Codex desktop update since launch on April 16. Not a version bump. A rewrite of what the app does.

Computer use on Mac. GPT-Image-1.5 inside the coding flow. An in-app browser that takes direct comments. Memory. And 90+ new plugins dropped in one release.

Weekly developer count jumped from 1.2M in January to 3M now. That's 150% growth in three months from a product that already owned the enterprise coding agent conversation.

Everybody's covering the feature list. Three things nobody's pointing at matter more.

Thing 1: Computer Use Is Background, Not Takeover

Read the headlines and you'd think Codex just seized your Mac. It didn't.

The computer use mode runs alongside you, not instead of you. OpenAI's own phrasing from the April 16 announcement: Codex can "take actions as directed in said applications, and, in the case of Mac users, even do so while you continue manually using your computer simultaneously to your agents working in the background."

That phrase matters. Anthropic's computer use, launched October 2024, requires you to hand over the mouse. Watching the cursor move by itself is jarring and unusable for real work. You go make coffee.

OpenAI flipped the model. Codex now does the Jira ticket update, the Slack thread dig, the screenshot annotation — in a sandbox layer — while your keyboard stays in Cursor or VS Code. You don't stop coding to ask it a question.

The practical impact: Codex is the first mainstream agent that feels like a coworker instead of a robot assistant.

Availability: Mac first. EU and UK users are locked out until OpenAI finishes a regional compliance pass. Windows support is "soon" with no date.

Thing 2: GPT-Image-1.5 Isn't About Pretty Pictures. It's About Closing the Design Loop.

The press angle on GPT-Image-1.5 is generation quality. Miss the point.

The real shift is workflow compression. Before this update, a frontend task looked like: take screenshot, open Figma, draft mockup, export, paste into chat, ask Codex to implement. Five windows, three apps, two copy-pastes.

Now it's: screenshot the bug, tell Codex "show me three redesigns in the same dimensions, then pick your favorite and patch the JSX." One conversation, no context switch.

Real iconography and precise brand colors remain the weakness — Stable Diffusion's last gen variants still beat it on 2D art from scratch. But for "make this card 10% taller and swap the accent color," it wins because it never leaves the editor.

Thing 3: The 90 Plugins Are a Trojan Horse for MCP

OpenAI called it "90+ additional plugins." Look closer. The release bundle has three categories mashed into one number: skills, app integrations, and MCP servers.

This is the first time a major AI vendor has shipped MCP servers as a first-class install experience. Click an integration. It registers. Done. No npm install, no JSON editing, no stdio plumbing.

The integration list reads like an enterprise wishlist: Atlassian Rovo for Jira and Confluence, CircleCI and GitLab Issues for CI/CD, Microsoft for Teams and Office.

For developers building on the Model Context Protocol, this is validation at a level the spec hasn't had before. GitHub's official MCP server added Streamable HTTP the same week. The stack is consolidating fast.

The sleeper feature buried in the announcement: the in-app browser now treats webpage comments as agent instructions. Highlight a button, type "this should be disabled when the form is invalid," and Codex reads it as a task. That's a UX primitive other agent tools will copy within six months.

What the Memory Feature Actually Does

Preview memory shipped alongside the big three. It's not ChatGPT-style trivia recall. It's a behavior model.

Codex now remembers your corrections. Tell it "I prefer tabs over spaces" once and it stops asking. Correct its import sort style twice and it internalizes the pattern for every future file.

The catch: memory is not available to Enterprise, Education, EU, or UK users yet. And unlike ChatGPT's memory, there's no per-project isolation yet.

Who This Actually Kills

Not Cursor. Cursor owns the "IDE with AI" category and this update doesn't invade it.

The real casualty is the middle layer: standalone agent apps that were trying to sit between your terminal and your ticketing system. Tools that marketed "autonomous engineer on your desktop" now have to explain why you'd use them when Codex is free with a ChatGPT subscription.

The 3M Weekly Developer Number

OpenAI confirmed 3M weekly developers use Codex. That's roughly 10% of the global professional developer population. GitHub Copilot reported about 10M paid seats in its last update. Codex is the free-tier version of that scale, running on ChatGPT Plus and Pro accounts.

The implication for hiring: "familiar with Codex" is now table stakes for any AI-forward engineering role. Expect it on job specs by July.

Frequently Asked Questions

What is the OpenAI Codex desktop app?
Codex is OpenAI's desktop coding agent for ChatGPT Plus and Pro subscribers, available on macOS and Windows. It runs an AI agent that can write code, browse your codebase, execute shell commands, and as of April 16, 2026 control other Mac apps, generate images, and use 90+ plugins.

How does Codex computer use compare to Anthropic's computer use?
Anthropic's version takes over your mouse and keyboard, so you can't work while it runs. Codex runs computer actions in the background while you keep using your machine.

How much does OpenAI Codex cost in April 2026?
Codex is included in ChatGPT Plus ($20/month) and ChatGPT Pro ($200/month). The 90+ plugins and computer use mode are included at no extra charge.

What are the best alternatives to OpenAI Codex in 2026?
The closest IDE-native alternative is Cursor. For agent-style coding, Claude Code and GitHub Copilot Workspace cover different slices. For visual app building, Lovable 2.0 handles full-stack generation from prompts.