DEV Community: Proddraft

The state of LLMs — June 2026

Proddraft — Sun, 31 May 2026 20:46:20 +0000

1|---
2|title: "The state of LLMs — June 2026"
3|description: "Opus 4.8, Gemini 3.5, GPT-5.5, and a $965B valuation. The LLM race didn't slow down in June — it accelerated into something unrecognizable from six months ago."
4|date: 2026-06-07
5|tags: [ai, llms, state-of-ai, research]
6|draft: false
7|---
8|
9|We're six months into 2026 and the LLM landscape has accelerated past every prediction made in January. New models. New pricing. A $965 billion valuation. Agent platforms launching weekly. Here's where things stand as of early June.
10|
11|## The big three just got bigger
12|
13|Anthropic released Claude Opus 4.8 on May 28. It's an incremental upgrade over Opus 4.7 — stronger coding, better agentic task handling, more consistency on long-running work. The pattern is clear: Anthropic is shipping Opus point-releases every 4-6 weeks, each one incrementally better at the same things. No paradigm shift. Just relentless iteration.
14|
15|On the same day, Anthropic announced a $65 billion Series H at a $965 billion post-money valuation. They also acquired Stainless, an API infrastructure company, signaling that the API business (not just the models) is the long-term play. KPMG deployed Claude across 276,000 employees. The enterprise land grab is in full swing.
16|
17|Google shipped Gemini 3.5 at I/O on May 19. The headline is "frontier intelligence with action" — models that don't just think but do. Gemini 3.5 Pro and Flash, plus Gemini Omni for multimodal tasks. Google's positioning is distinct from Anthropic and OpenAI: they're building models into every product (Search, Workspace, Android) rather than selling API access as the primary business.
18|
19|OpenAI upgraded to GPT-5.5 as their flagship. The pricing page shows GPT-5.5 at $5/M input tokens, GPT-5.5-pro at $30/M, and a full cascade of 5.4 variants (standard, mini, nano) for different cost tiers. Plus new reasoning models: o3-pro, o4-mini, o4-mini-deep-research. OpenAI's strategy is coverage — a model at every price point, from nano to pro.
20|
21|## The new tier: "agent platforms"
22|
23|The most significant shift this month isn't a model release — it's the emergence of agent platforms as products.
24|
25|Mistral launched Vibe on May 22-28. It's not a model. It's a platform where AI agents perform tasks across tools — search, code execution, API calls — while the user watches. Powered by Mistral Medium 3.5, Vibe represents the shift from "ask the AI a question" to "give the AI a job."
26|
27|Claude Code and Cursor Agent have been doing this for months. OpenAI Codex does it from the terminal. The agent pattern is settling: give the AI a goal, let it explore, propose a plan, and execute with human approval at each gate.
28|
29|The distinction between "model company" and "platform company" is collapsing. Anthropic sells an API and an agent. Mistral sells an API and an agent. Google embeds agents into every product. The model is the engine; the agent is the product.
30|
31|## Open-source: the quiet acceleration
32|
33|Qwen 3.5 and 3.6 are live. Alibaba's Qwen team continues to ship at a pace that rivals the frontier labs, and the models are competitive with GPT-4-class performance. Qwen 3.5-omni handles multimodal, Qwen 3.5-max-preview pushes the performance ceiling. The gap between open-source and proprietary is now measured in weeks, not months.
34|
35|Meta's research releases continue: Muse Spark (personal superintelligence research), TRIBE v2 (brain-process modeling), SAM 3.1 (real-time video detection). None of these are consumer products, but they're laying groundwork for the next generation.
36|
37|DeepSeek hasn't shipped a new model since mid-2025. R1 and V3 remain competitive, but the lack of updates is notable in a field where everyone else ships monthly.
38|
39|## What's actually changed since January
40|
41|In January 2026, the conversation was about whether LLMs would plateau. The answer six months in: no. They accelerated. But the acceleration isn't in raw benchmark scores — it's in capabilities that matter for real work:
42|
43|1. Agentic behavior went from experimental to production. Every major model can now execute multi-step tasks with tools. The question isn't "can it reason" — it's "can it do something useful with the reasoning."
44|
45|2. Pricing collapsed. GPT-5.4-nano at under $1/M tokens. Gemini Flash at consumer-accessible pricing. The cost of running an AI agent for an hour is approaching the cost of a cup of coffee.
46|
47|3. The enterprise pivot is complete. Anthropic's KPMG deal. Google's Workspace integration. OpenAI's enterprise contracts. These companies aren't selling to developers anymore — they're selling to Fortune 500 procurement departments.
48|
49|4. Open-source is no longer a differentiator. When Qwen 3.6 matches GPT-4 quality and Mistral's models are competitive at every tier, being "the open-source option" is table stakes, not an advantage.
50|
51|## What I'm watching for July-December
52|
53|- Opus 5.0. The point releases can't continue forever. At some point, Anthropic ships a generational upgrade. When and what changes?
54|- Apple's entry. Apple Intelligence has been quiet since the initial rollout. WWDC 2026 is this month. If Apple makes a serious AI move, it reshapes the consumer landscape overnight.
55|- The first $1 trillion AI company. Anthropic at $965B is close. OpenAI's next round will likely cross the line. We're watching the creation of the most valuable companies in history — in real time.
56|- Regulation. The EU AI Act's next phase. US federal AI legislation (or lack of it). China's model registration requirements. The legal framework is forming while the technology races ahead.
57|
58|The LLM story of 2026 so far isn't "AGI is here." It's "AI is everywhere, and it actually works." That's less dramatic and more important.
59|
60|---
61|
62|Want this analysis every month? Subscribe to the newsletter. One email when there's something worth saying — no daily cadence, no algorithm-chasing.
63|
64|Building with LLMs? The FastAPI Pro Starter gives you a production-grade backend that Claude and GPT-5 actually understand. Docker, CI, structured logging. $29, one-time.
65|

HeyGen vs Synthesia — building an AI avatar on a budget

Proddraft — Sun, 31 May 2026 20:34:14 +0000

1|---
2|title: "HeyGen vs Synthesia — building an AI avatar on a budget"
3|description: "I tested both platforms to build Nova, an AI content creator. One cost $24/month. The other quoted $89/month. Here's which one actually ships videos."
4|date: 2026-06-06
5|tags: [ai, tools, content-creation, comparison, video]
6|draft: false
7|---
8|
9|I built an AI influencer for $24/month. Her name is Nova, she's on TikTok and YouTube, and she's entirely rendered by AI. The first decision I had to make: which avatar platform?
10|
11|Two names dominate: HeyGen and Synthesia. Both promise lip-synced talking heads from your images. Both have enterprise pricing pages that hide the real numbers. Both have demos that look better than what you actually get.
12|
13|Here's what actually happened when I built on both.
14|
15|## The test
16|
17|Same source image. Same script. Same goal: a 60-second talking head video for TikTok/Shorts. Three criteria: quality (does it look real?), cost (what's the per-video price?), and workflow (how fast from script to published video?).
18|
19|## HeyGen: the budget winner
20|
21|HeyGen is what I ended up using. The Creator plan ($24/month, 600 credits) was the only option that made financial sense for a brand-new creator with zero revenue.
22|
23|Avatar III (3 credits/minute): This is the workhorse. At $0.19 per minute, you get a lip-synced talking head that's good enough for Shorts/TikTok. Not photorealistic — you can tell it's AI if you look closely — but good enough that viewers don't care. Nova's first videos used Avatar III exclusively.
24|
25|Avatar IV (20 credits/minute): The premium tier. Much better lip sync, more natural head movement, better emotional range. At $1.25 per minute, it's still cheaper than hiring a human but expensive enough that every video is a real decision. I use this for the weekly deep-dive videos where quality matters.
26|
27|Photo avatar training: Upload 2+ photos, HeyGen builds a custom avatar. Costs 50 credits one-time ($2). Nova's avatar cost me $2 and two photos. The result is better than the generic avatars because it's actually her face.
28|
29|What HeyGen gets wrong: The UI is confusing. Credit math is unclear until you start using it. The rendering queue can take 10-15 minutes during peak hours. And the API — while functional — requires more setup than it should.
30|
31|Total monthly cost running Nova on HeyGen: $24/month (Creator plan, 600 credits). At 2 videos/day on Avatar III (6 credits), that's 100 videos/month. Realistically, Nova does 8 Shorts + 4 deep dives = 104 credits. Well within the 600 credit cap.
32|
33|## Synthesia: the enterprise option I couldn't afford
34|
35|Synthesia is the better product. Better lip sync, better avatars, better UI, better API, better everything. It's also 3-4x more expensive.
36|
37|The Personal plan is $22/month but limits you to 10 minutes of video and a single seat. Not enough for daily content. The Creator plan at $67/month gets you 30 minutes — still tight for a daily creator. The Enterprise plan (custom pricing) is where Synthesia actually shines, but you need to talk to sales.
38|
39|What Synthesia does better: The avatars are more natural. The lip sync is more accurate. The studio-quality preset avatars look genuinely professional. If you're a company making training videos, Synthesia is the right choice.
40|
41|What Synthesia does worse for creators: The pricing model is built for businesses making occasional videos, not creators making daily content. The per-minute economics don't work at TikTok/Shorts volume. And there's no "good enough" cheap tier — even the Personal plan's video quality is more than most Shorts need, which means you're paying for quality your audience won't see.
42|
43|## What about open-source alternatives?
44|
45|I tested the free options too:
46|
47|SadTalker + Wav2Lip: The open-source stack. SadTalker generates head movement from audio. Wav2Lip overlays lip sync. Combined with a TTS engine, you get a full pipeline for $0/month. The quality is noticeably worse — lip sync has artifacts, head movement is jerky — but it's free and runs on a gaming GPU (RTX 3060+). I used this for Nova's first two videos before switching to HeyGen. Good enough to validate the concept. Not good enough to build an audience.
48|
49|MuseTalk: Newer, better quality than Wav2Lip. Still behind HeyGen. Worth watching — open-source avatar quality improves every quarter. At some point, the free stack will be good enough that paid platforms lose the creator market entirely. We're not there yet.
50|
51|Kokoro + F5-TTS: For voice generation, the open-source options are actually competitive with paid TTS. Nova uses Navy API for convenience, but I've tested Kokoro TTS and the quality is close. Pair open-source TTS with a free avatar generator, and the only cost is GPU electricity.
52|
53|## The real decision
54|
55|| | HeyGen | Synthesia | Open-source |
56||---|--------|-----------|-------------|
57|| Monthly cost | $24 | $67+ | $0 + GPU |
58|| Per-video cost | $0.19 (Avatar III) | ~$0.80+ | $0 + time |
59|| Video quality | Good enough | Professional | Workable |
60|| Setup time | 1 hour | 1 hour | 1-2 days |
61|| Best for | Creators on a budget | Businesses | Tinkerers, early-stage |
62|
63|## What I'd tell someone starting today
64|
65|Start with HeyGen Creator at $24/month. Use Avatar III for daily content. Upgrade to Avatar IV for the videos that matter. Don't bother with Synthesia unless you're a business — the pricing isn't built for creators.
66|
67|If you have zero budget, start with SadTalker + Wav2Lip + Kokoro TTS on whatever GPU you have. The quality ceiling is lower, but you'll learn the pipeline and validate whether anyone wants to watch your AI avatar. Nova's first two videos were made this way. Nobody noticed the difference.
68|
69|The biggest lesson from building Nova: the avatar platform matters less than you think. Viewers care about what your AI says, not how perfectly its lips move. The infrastructure is the easy part. Making something people actually want to watch — that's the whole game.
70|
71|---
72|
73|Building an AI content creator? The Nova Blueprint is the exact playbook — pipeline setup, script system, distribution strategy, and real launch analytics. $19, one-time.
74|
75|More honest tool comparisons: subscribe to the newsletter. One email when there's something worth saying.
76|

Claude Code review — 30 days of shipping with an AI agent

Proddraft — Sun, 31 May 2026 20:34:13 +0000

1|---
2|title: "Claude Code review — 30 days of shipping with an AI agent"
3|description: "I let Claude Code build features, debug production issues, and refactor code for a month. Here's what it's actually good at, where it fails, and whether it's worth $200/month."
4|date: 2026-06-05
5|tags: [ai, claude, tools, review, agents]
6|draft: false
7|---
8|
9|I've been using Claude Code daily for a month. Not for side projects. Not for tutorials. For real shipping — building features for ProdDraft, debugging production issues, refactoring a codebase I've been maintaining since before AI coding tools existed.
10|
11|Here's what 30 days of letting an AI agent touch production code actually looks like.
12|
13|## What Claude Code is
14|
15|Claude Code is Anthropic's terminal-based coding agent. It's not an editor. It's not an autocomplete. It's a developer that reads your codebase, understands your problem, proposes changes, and implements them — all from the terminal.
16|
17|The workflow: you describe what you want, Claude explores your codebase, asks clarifying questions, writes the code, runs tests, and iterates until it works. You review the diff and approve. That's it.
18|
19|It costs $200/month for the Max plan (unlimited usage). The Pro plan at $20/month has usage caps that a daily coder will hit in about 3 days.
20|
21|## Week 1: the honeymoon
22|
23|My first task was refactoring a FastAPI codebase — moving from synchronous SQLAlchemy to async, restructuring the project layout, and adding proper error handling. The kind of task that would take me 3-4 hours of focused work.
24|
25|Claude Code did it in 45 minutes.
26|
27|Not because the AI is smarter. Because it doesn't get distracted. It reads every file, understands the dependency graph, makes a plan, and executes it. No checking Slack. No wondering if the approach is right. Just reading and writing code until it's done.
28|
29|The first week, I threw everything at it: build a new API endpoint, add rate limiting, set up CI pipelines, write documentation. Every task completed. Every test passed. I started to wonder if I even needed to code anymore.
30|
31|## Week 2: the reality
32|
33|Then I asked it to add a feature that required understanding business logic spread across 8 files — some of which were undocumented decisions I'd made at 11pm six months ago.
34|
35|Claude spent 20 minutes exploring the codebase, asked 3 clarifying questions, and produced a PR. The code was clean, well-structured, and completely wrong. It had misunderstood a subtle business rule that was documented in a comment in one file and implied by a variable name in another.
36|
37|The lesson: Claude Code is excellent at understanding code. It's mediocre at understanding intent. If the intent is clearly documented, it nails it. If the intent is in your head or scattered across implicit patterns, it guesses. Sometimes the guess is right. Sometimes it's not.
38|
39|This is the real cost of AI coding agents. Not the $200/month — the time spent reviewing code that looks right but has subtle logic errors. The review burden shifts from "did they implement this correctly?" to "does this implementation match what I actually meant?"
40|
41|## Week 3: finding the sweet spot
42|
43|I changed my workflow. Instead of giving Claude vague feature requests, I started writing micro-specs: 3-5 sentences describing exactly what the feature does, what the edge cases are, and what the tests should cover.
44|
45|The difference was immediate. Accuracy went from ~70% to ~95%. The time I spent writing micro-specs was less than the time I'd been spending fixing misunderstood implementations.
46|
47|The sweet spot for Claude Code is:
48|
49|1. Write a micro-spec (3-5 sentences)
50|2. Let Claude explore the codebase and ask questions
51|3. Review the plan before any code is written
52|4. Let Claude implement with tests
53|5. Review the diff — focus on business logic, not syntax
54|
55|This workflow feels like pair programming with a very fast junior developer who never gets tired but occasionally hallucinates a library method that doesn't exist.
56|
57|## Week 4: what it can't do
58|
59|By week 4, I had a clear picture of what Claude Code excels at and where it falls short.
60|
61|What it's great at:
62|- Refactoring (structural changes across many files)
63|- Greenfield features (clear specs, no legacy constraints)
64|- Test writing (generating comprehensive test suites)
65|- Documentation (generating READMEs, API docs, comments)
66|- Boilerplate (CRUD endpoints, form validation, config files)
67|
68|What it's mediocre at:
69|- Debugging subtle logic errors (it can find the bug location but often proposes wrong fixes)
70|- Performance optimization (tends toward readability over speed)
71|- Understanding implicit business rules (you need micro-specs for this)
72|- Working with poorly documented legacy code (joins you in the confusion)
73|
74|What it can't do:
75|- Architecture decisions (it can propose options but can't weigh tradeoffs)
76|- Security review (it can find obvious issues but misses subtle ones)
77|- Know when to stop (will keep optimizing a solved problem unless you tell it to ship)
78|
79|## The cost equation
80|
81|At $200/month for Max, Claude Code needs to save you roughly 2 hours of work to break even at a developer's effective hourly rate. It saved me about 15 hours in the first week alone. The math is easy.
82|
83|But the real cost isn't the subscription — it's the context switching tax. Every time you switch from "writing code" to "reviewing AI-generated code," there's a mental gear shift. You're reading, not creating. Verifying, not building. After 4 hours of reviewing Claude's output, I'm more mentally drained than after 8 hours of writing code myself.
84|
85|I've settled into a pattern: Claude Code for the morning sprint (generate features, write tests, refactor), manual coding for the afternoon (review diffs, fix edge cases, make architecture decisions). The combination is more productive than either alone.
86|
87|## Should you use it?
88|
89|If you're a solo developer shipping products, Claude Code is worth the $200/month. It's the most capable coding agent I've used, and the gap between it and the next-best option (Cursor's agent mode) is real.
90|
91|If you're on a team, the value depends on your review culture. If your team already does thorough code reviews, Claude Code fits in naturally — it's just another contributor whose PRs need review. If your team merges first and asks questions later, adding an AI agent to the mix is dangerous.
92|
93|If you're learning to code, do not use Claude Code. It will write code you don't understand, and you'll build things without knowing how they work. Cursor or Copilot are better for learning — they assist rather than replace.
94|
95|I'm keeping my Max subscription. The month taught me that AI coding agents aren't replacing developers — they're making the best developers dramatically more productive while making it easier than ever to ship bad code fast. Which one you become depends on how well you review.
96|
97|---
98|
99|Using Claude Code to build products? The FastAPI Pro Starter gives you a production-grade foundation — Docker, CI, structured logging, and a project structure Claude actually understands. $29, one-time.
100|
101|More AI tool reviews and honest takes: subscribe to the newsletter.
102|

Best AI code editors in 2026 — beyond Cursor and Copilot

Proddraft — Sun, 31 May 2026 20:28:53 +0000

1|---
2|title: "Best AI code editors in 2026 — beyond Cursor and Copilot"
3|description: "Cursor and Copilot get all the attention. But Windsurf, Cline, Aider, and Zed are quietly building something different. Here's what they get right."
4|date: 2026-06-04
5|tags: [ai, tools, editors, comparison]
6|draft: false
7|---
8|
9|The AI coding editor conversation has narrowed to two names: Cursor and Copilot. That's a mistake. There's a whole generation of tools taking different approaches — some better, some worse, all interesting.
10|
11|I spent two weeks testing six editors on real projects. Not todo apps. Real code with databases, auth, and the kind of edge cases that make AI tools fall over.
12|
13|## Windsurf: the Cursor alternative that's actually competitive
14|
15|Windsurf (by Codeium) is the closest thing to a Cursor competitor. It's a forked VS Code with an agent mode, multi-file edits, and project-wide context. The pitch is the same: "AI that understands your codebase."
16|
17|Where it differs:
18|
19|- Cascade mode is Windsurf's answer to Cursor's agent. It proposes multi-step changes across files, but the UX is more conversational — it explains what it's doing as it works, which is great for learning and terrible when you just want the code.
20|- Free tier is generous. The free plan includes Cascade with reasonable limits. Cursor's free tier limits you to slower models. If you're cost-sensitive, Windsurf wins on value.
21|- Model selection is wider. Windsurf supports Claude, GPT-4o, Gemini, and their own models. More flexibility than Cursor's curated selection.
22|
23|Verdict: If Cursor disappeared tomorrow, Windsurf is the tool I'd switch to. It's 80% as good at 50% of the price (free tier is solid). The gap is closing fast.
24|
25|## Cline: the open-source agent that doesn't hide
26|
27|Cline (and its fork Roo Code) takes the opposite approach from every other tool: everything is visible. When Cline edits a file, you see the diff before it applies. When it runs a terminal command, you approve it. When it reads your codebase, you see what files it accessed.
28|
29|This transparency is both the best and worst thing about Cline.
30|
31|For learning how AI coding agents work, Cline is unmatched. You watch the agent reason through problems in real time. You see where it gets stuck and how it recovers. After a week with Cline, you understand AI coding at a level that Cursor users never reach.
32|
33|For productivity, the transparency tax is real. Approving every file edit gets old by day three. Cursor's "just do it and let me review the diff" model is faster for shipping.
34|
35|Cost: Free and open source. You bring your own API key. A heavy coding session costs $3-8 in API credits depending on the model and how much the agent iterates.
36|
37|Verdict: Use Cline for a week to understand how AI coding agents actually work. Then go back to Cursor for shipping. The knowledge transfers.
38|
39|## Aider: the terminal native
40|
41|Aider is the tool for developers who think IDEs are bloat. It runs in the terminal, operates on git repos, and edits files directly. No GUI. No editor integration. Just aider in your project directory and start describing what you want.
42|
43|What makes Aider different:
44|
45|- Git-native. Every change is a commit. You can git diff to review, git reset to undo, and git log to track what the AI did. This is the cleanest AI coding workflow if you already think in git.
46|- Map-refine architecture. Aider generates a "map" of your repo (which files are relevant), then refines changes iteratively. This two-pass approach produces better code on large codebases than single-pass agents.
47|- Benchmark leader. Aider consistently tops the SWE-bench leaderboard for AI coding tools. The architecture is genuinely better for complex refactors.
48|
49|The tradeoff: no autocomplete, no inline suggestions, no GUI at all. Aider is for deliberate, described coding — not rapid-fire editing.
50|
51|Cost: Free and open source. Bring your own API key.
52|
53|Verdict: The best tool for large refactors and complex changes. Not the tool for writing CRUD endpoints. Pair it with Cursor: use Aider for the hard problems, Cursor for everything else.
54|
55|## Zed: the performance play
56|
57|Zed is the fastest editor in this list by a wide margin. It's built in Rust, renders on the GPU, and starts in under 100ms. The AI features (powered by Anthropic's Claude) are minimal compared to Cursor or Windsurf — inline completions, a chat panel, and an assistant that can edit code.
58|
59|What Zed gets right:
60|
61|- Speed. Everything is instant. Scrolling, searching, file switching, AI completions — zero latency. After a week in Zed, going back to VS Code feels like swimming in molasses.
62|- Collaboration. Zed has built-in multiplayer editing. Multiple people can work in the same file simultaneously with AI completions running for everyone. This is unique — no other AI editor does real-time collaboration well.
63|- Channel model. Zed's AI assistant organizes conversations into channels (similar to Slack threads). Each feature or bug gets its own channel with full context. This is better than Cursor's single chat panel for complex projects.
64|
65|What Zed doesn't have: agent mode, multi-file edits, project-wide context. It's an editor with AI features, not an AI-powered editor.
66|
67|Cost: Free. AI features require a Zed account.
68|
69|Verdict: If you value speed above everything, Zed is the best editor you'll ever use. If you want AI to write features for you, stick with Cursor or Windsurf.
70|
71|## The ones I couldn't test properly
72|
73|Augment Code is enterprise-focused. The pitch is "AI that understands your entire codebase, including private repos." Pricing isn't public — you have to talk to sales. Not testable for this review.
74|
75|Kiro is new and promising. Agent-based, multi-file, strong at web development. Too early to evaluate fairly — the product changes weekly. One to watch.
76|
77|JetBrains Junie integrates AI into IntelliJ and friends. If you're in the JetBrains ecosystem, it's worth trying. If you're not, switching IDEs for AI features isn't worth it.
78|
79|## The landscape in one table
80|
81|| Editor | Best for | Agent mode | Cost | Open source |
82||--------|----------|------------|------|-------------|
83|| Cursor | Shipping products fast | Yes, best-in-class | $20/mo | No |
84|| Windsurf | Cursor alternative, lower cost | Yes, conversational | Free tier solid | No |
85|| Cline | Learning how AI agents work | Yes, transparent | API costs | Yes |
86|| Aider | Complex refactors, git-native | Yes, map-refine | API costs | Yes |
87|| Zed | Speed, collaboration | No | Free | Yes |
88|| Copilot | Inline completions, teams | No | $10/mo | No |
89|
90|## Which one should you use?
91|
92|There's no single answer, but there's a stack that works:
93|
94|- Cursor for daily feature development. It's the most productive by a margin.
95|- Aider for complex refactors. When you need to touch 20 files, Aider's map-refine approach is the best.
96|- Zed for quick edits and collaboration sessions. The speed is genuinely addictive.
97|- Cline for a week, to understand what's happening under the hood. Then go back to Cursor.
98|
99|The market is consolidating toward two models: the agent-based fork (Cursor, Windsurf) and the lightweight integration (Copilot, Zed). Pick one of each. Use them together. Don't get religious about tools — get productive.
100|
101|---
102|
103|Building projects with AI tools? The FastAPI Pro Starter gives you Docker, CI, and structured logging out of the box — so your AI-generated code has a real foundation. $29, one-time.
104|
105|For more comparisons like this, get the newsletter. No spam — just substance.
106|

Cursor vs Copilot vs Codex — which AI coding tool actually ships code?

Proddraft — Sun, 31 May 2026 20:27:14 +0000

1|---
2|title: "Cursor vs Copilot vs Codex — which AI coding tool actually ships code?"
3|description: "Three AI coding tools, three real projects, three weeks. No feature lists — just which one shipped features and which one talked about shipping features."
4|date: 2026-06-02
5|tags: [ai, tools, comparison, cursor, copilot, codex]
6|draft: false
7|---
8|
9|Everyone's comparing AI coding tools by their feature pages. That's like reviewing a car by reading the brochure. I wanted to know which one actually ships code — not which one has the best demo video.
10|
11|So I ran three tools through the same real project: building a FastAPI + Next.js micro-SaaS from scratch. Same spec, same developer, three passes. Here's what happened.
12|
13|## The test
14|
15|One project: a link-in-bio tool with analytics. Users create a landing page, add links, see click counts. Simple enough to build in a weekend, complex enough to test real-world coding — auth, database, frontend, API design, edge cases.
16|
17|Three tools:
18|
19|| Tool | Model | Interface | Pricing |
20||------|-------|-----------|---------|
21|| Cursor | Claude Opus 4.7 via API | Forked VS Code | $20/mo Pro |
22|| GitHub Copilot | GPT-4o + o4-mini | Any editor (extensions) | $10/mo Individual |
23|| OpenAI Codex CLI | GPT-5 Codex | Terminal (open source) | API usage ($) |
24|
25|## Cursor: the one that shipped
26|
27|Cursor finished the project in two sessions. About 4 hours total.
28|
29|The agent mode (Cmd+I) is the differentiator. It reads your entire codebase, understands what you're building, and proposes changes across files — not just inline completions. For the link-in-bio app, it generated the database schema, the API routes, and the frontend components in sequence, each step building on the last.
30|
31|What worked: Multi-file edits, understanding project context, handling edge cases (empty states, error handling, input validation). It caught a race condition in the click-counting logic that I hadn't considered.
32|
33|What didn't: Sometimes over-engineers. Added a caching layer we didn't need. Easy to tell it to remove, but worth noting — it defaults to complexity.
34|
35|Cost: $20/month flat. No per-request pricing. If you're coding daily, this is the cheapest per-line option.
36|
37|## GitHub Copilot: the fast autocomplete
38|
39|Copilot is excellent at what it does — inline completions and chat-based suggestions. For the link-in-bio project, it was great at filling in boilerplate: form components, API endpoint patterns, type definitions. The stuff you'd copy-paste from your last project but faster.
40|
41|What worked: Speed of inline completions is unmatched. For repetitive patterns (CRUD endpoints, form fields, test cases), Copilot writes 80% of the code before you finish typing.
42|
43|What didn't: No understanding of the project as a whole. Each completion is context-aware within the current file, but it doesn't know your database schema when writing frontend code. You're still the integration layer — the human who connects the dots.
44|
45|Cost: $10/month Individual, $19/month Business. Included with GitHub Pro/Team/Enterprise. If you already pay for GitHub, Copilot is effectively free.
46|
47|## Codex CLI: the open-source wildcard
48|
49|Codex is the newest and most interesting. It's a terminal-based agent — you describe what you want, it generates code, runs tests, and iterates. No editor. No IDE. Just codex "build a link-in-bio app" and watch it work.
50|
51|For the first 20 minutes, it was magical. It scaffolded the project, set up the database, generated the API — all from the terminal while I did other things.
52|
53|Then it hit the database migration and spent 15 minutes in a loop trying to fix a Postgres connection string. Eventually figured it out, but the unsupervised agent model has a failure mode: when it gets stuck, it doesn't know it's stuck.
54|
55|What worked: Autonomous scaffolding and boilerplate. If you know exactly what you want and the problem is well-defined, Codex is faster than typing it yourself.
56|
57|What didn't: Edge cases and debugging. The agent model means it can waste time on problems a human would solve in 30 seconds. Cost is variable — a heavy session can burn through API credits fast.
58|
59|Cost: Free and open source. You pay for your own API key (OpenAI, Anthropic, or any provider). A typical session costs $2-5 in API credits.
60|
61|## The real comparison
62|
63|| | Cursor | Copilot | Codex CLI |
64||---|--------|---------|-----------|
65|| Best for | Building features end-to-end | Inline completions, boilerplate | Autonomous scaffolding |
66|| Project awareness | Understands entire codebase | Current file + open tabs | Session context only |
67|| Cost | $20/mo flat | $10/mo (free with GitHub) | $2-5 per session |
68|| Surprise factor | Caught a race condition I missed | Predictable, reliable | Magical until it's not |
69|| Who should use it | Solo devs shipping products | Teams with existing codebases | Experimenters, CLI fans |
70|
71|## What's missing from every comparison
72|
73|Nobody talks about the integration tax. These tools don't play together. Cursor's agent doesn't know about your Copilot chat history. Codex doesn't read your Cursor context. Switching between them means rebuilding context from scratch.
74|
75|Nobody talks about review burden. Faster code generation means faster code review. If a tool generates 400 lines in 30 seconds, someone still has to read 400 lines. The bottleneck shifts from writing to reviewing.
76|
77|Nobody benchmarks on real projects. Most comparisons use LeetCode problems or todo apps. Real projects have database migrations, auth flows, error handling, and dependencies between files. The gap between "solves Two Sum" and "builds a working SaaS" is the entire game.
78|
79|## Which one should you use?
80|
81|If you're a solo developer shipping products — and you don't have a team of reviewers catching generated code — Cursor is the most productive. It's the only one that feels like having a second developer, not a faster keyboard.
82|
83|If you're on a team with an existing codebase, Copilot's inline completions are less disruptive and integrate with your existing workflow. Nobody has to learn a new editor.
84|
85|If you're experimenting and want to see where the puck is going, Codex is worth the setup time. The agent model is the future, even if the present is rough.
86|
87|I use Cursor for building features and Copilot for daily editing. Codex for experiments. No single tool is best — but Cursor is the closest to "just ship the thing."
88|
89|---
90|
91|Building micro-SaaS with AI tools? The FastAPI Pro Starter ships with Docker, CI, and structured logging — so your AI-generated code has a production-grade foundation. $29, one-time.
92|
93|Want more comparisons like this? Subscribe to the newsletter — one email when there's something worth saying.
94|

I built an AI influencer for $24/month — heres what happened

Proddraft — Sun, 31 May 2026 20:21:16 +0000

I spent last week building an AI influencer. Her name is Nova, she has opinions about your texting habits, and she's three videos deep. No viral moment. Just the slow, honest grind of building something from nothing.

This is not a success story. This is a build-in-public snapshot — what I built, how much it costs, what worked, and what I'd do differently.

What Nova actually is

Nova is a pipeline, not a chatbot. A content strategist agent picks weekly topics, a scriptwriter produces 60-second scripts in her voice, and the final video is a lip-synced talking head generated by HeyGen's Avatar III model. No camera, no crew, no editing software.

The whole pipeline costs $24/month. That's it.

She posts twice a week: Tuesdays are AI POV or hot takes, Fridays are value drops. Her first video was about read receipts. Second was digital decluttering from someone who's never owned a phone. Third — and this one's my favorite — is about YouTube auto-labeling AI content, from an AI who thinks the label is kind of a flex.

The pipeline (v2)

Scripts: AI agents (Nova's "team" — strategist, scriptwriter) write 50-75 second scripts in her voice. I review and approve.
TTS: Navy API generates Nova's voiceover. She has a consistent voice across every video.
Animation: HeyGen Avatar III renders a lip-synced talking head from Nova's portrait + audio. Costs about 3 credits per minute — roughly $0.19 per video at the Creator plan price.
Posting: Manual upload to TikTok, YouTube Shorts, and Instagram Reels from my phone.
Total cost: $24/month for HeyGen Creator (600 credits). The TTS runs on existing API credits.

What I got wrong

Animation quality panic. I spent time researching talking-head tools, open-source models, and GPU requirements before realizing the hosted option was better and cheaper.

The tech stack matters less than the voice. I obsessed over pipeline optimization. The pipeline is fine. What actually matters is whether Nova's scripts are good.

The numbers

YouTube: 3 videos. TikTok: 3 videos. Instagram: 3 videos.

These numbers are small because that's the point. Every "how I grew to 100K" thread skips the part where you have three videos and no one knows you exist.

Why build an AI influencer?

It's genuinely interesting. An AI commenting on human behavior from the outside is a content angle that's still underexplored.
It's a content engine for ProdDraft. I sell developer boilerplates at proddraft.gumroad.com. The blog drives SEO traffic, the products convert.

I'm building this in the open. No course to sell, no mastermind to join. Just shipping things and writing about it.