The AI subscription grey market is now a GitHub trending category

5 min read 3 sources clear_take
├── "These repos are a coordinated grey-market response to AI vendor pricing"
│  └── top10.dev editorial (top10.dev) → read below

The editorial argues that three unrelated trending repos converging at once is not coincidence but signal. Together they represent developers systematically arbitraging the gap between flat subscription tiers and metered API pricing, driven by Claude Max windows tightening and o-series token costs becoming unpredictable.

├── "Turn a ChatGPT Plus subscription into an OpenAI-compatible API endpoint"
│  └── akihitohyh (GitHub, 244 pts) → read

The sub2api project exposes an authenticated ChatGPT web session as a drop-in /v1/chat/completions endpoint. The implicit position is that a $20/month subscription should be usable as production API infrastructure, sidestepping metered per-token billing entirely.

├── "Claude's pricing and usage limits deserve open pushback"
│  └── LinXiaoTao (GitHub, 593 pts) → read

The repository's confrontational name (FuckClaude) frames Anthropic's Max plan windows and pricing structure as user-hostile enough to warrant a tooling response. It's the loudest signal that developers running real workloads feel squeezed by the current subscription model.

└── "Reusable skill libraries are the legitimate way to reduce AI spend"
  └── dzhng (GitHub, 439 pts) → read

The skills repo takes a non-confrontational approach: curated Claude Skills and reusable agent workflows that compress the number of turns needed to get useful output. The position is that better prompt scaffolding — not vendor circumvention — is the sustainable way to lower cost per useful result.

What happened

Three repositories trended on GitHub this week, and on the surface they look unrelated. LinXiaoTao/FuckClaude climbed to a score of 593. dzhng/skills, a curated collection of Claude Skills and reusable agent workflows, hit 439. akihitohyh/chatgpt-register-sub2api — a tool that turns a ChatGPT web subscription into an OpenAI-compatible API endpoint — pulled 244. Different authors, different stacks, different framings.

Read them together and the pattern is obvious. All three are answers to the same question: how do I get more value out of an AI vendor's pricing model than the vendor intended? FuckClaude's name is the loudest signal, but the sub2api project is the more instructive artifact. It takes an authenticated ChatGPT Plus session and exposes it as an OpenAI-compatible `/v1/chat/completions` endpoint, so a $20/month subscription becomes drop-in replacement infrastructure for what would otherwise be metered API spend. The skills repo is less confrontational but sits on the same axis: reusable prompt scaffolding designed to compress the number of Claude turns you burn to get useful output.

None of these are new ideas in isolation. Reverse-engineered ChatGPT wrappers have existed since 2023. Skill libraries are as old as prompt engineering. What's new is the density: three of them in the trending list at the same time, from unrelated authors, aimed at two different vendors.

Why it matters

The AI infrastructure story of 2024 was capability. The story of 2025 has quietly turned into cost. Anthropic and OpenAI both raised or restructured pricing this year. Claude's Max plans introduced tighter usage windows. OpenAI's o-series reasoning models made per-token costs harder to predict. And developers running production workloads on top of these APIs are watching their COGS creep up in ways that the flat subscription tiers do not.

The grey-market response is what you'd expect from any market where retail pricing and wholesale pricing diverge by an order of magnitude. A ChatGPT Plus subscription costs $20/month for effectively unlimited GPT-4-class access under rate limits. The equivalent API usage — call it 20M tokens of output — would run you a few hundred dollars. Somebody was always going to build the arbitrage tool. Somebody built it, called it sub2api, and 244 people starred it in a week.

The vendor response so far has been fingerprinting, session-binding, and Terms of Service language. None of it has stopped the tools from shipping. Anthropic and OpenAI both explicitly prohibit programmatic access via consumer subscriptions in their ToS, so any team pulling one of these repos into their stack is picking up real contractual and IP exposure. That's a very different risk profile from `npm install openai`.

The skills repo is the interesting middle ground. It doesn't violate anyone's ToS — it just makes each API call do more work. That's the same economic pressure as sub2api, expressed legitimately: if the vendor won't cut the price, cut the number of calls. It's also the direction Anthropic itself is nudging developers with the Skills primitive: package reusable capability, reduce round trips, keep more logic client-side. When your customers' cost-optimization strategy and your vendor's product roadmap point at the same thing, you're probably right about where the market is going.

Community reaction on the repos themselves is instructive. The sub2api README has issues open from developers asking about proxying through residential IPs and dodging Cloudflare challenges — the usual scraper cat-and-mouse. FuckClaude's discussions read more like a support forum for people who blew through their Max plan window and want their workflow back. That's not a hacker crowd. That's a working-developer crowd that has decided the sticker price doesn't match the value they're getting.

What this means for your stack

Don't ship any of these into production. That part should be obvious, but it's worth stating: a subscription-wrapped API endpoint can be revoked, rate-limited, or fingerprinted out of existence on any given Tuesday, and if it takes your prod inference layer with it, that is a self-inflicted incident. The ToS exposure is real. So is the reliability profile.

What is worth doing is auditing where your team's AI spend is actually going, because the reason these repos are trending is that the answer is often "more than we expected." If you're running a Claude Code or Cursor workflow inside an engineering org, someone on your team has already priced out the delta between per-seat subscriptions and pay-as-you-go API calls, and they've noticed. If they haven't shown you the sub2api repo yet, they will.

The more durable takeaway is the skills-repo one. Bundling common workflows into reusable, tool-augmented skills — Anthropic's own primitive, or dzhng's community version, or your own internal library — is the legitimate version of the same cost pressure. Every skill you package is turns you don't have to burn. If you're not already treating agent skill libraries as a first-class piece of your AI stack, this is the quarter to start. Cheaper than negotiating enterprise contracts, cleaner than building a proxy layer, and it doesn't put your CTO on a call with Anthropic's legal team.

On the vendor side, expect pricing to move. When the arbitrage gets this visible, one of two things happens: enforcement tightens (session fingerprinting, hardware attestation, aggressive account bans) or pricing converges (subscription tiers that include real API quota, volume discounts that make the grey market not worth the ToS risk). Anthropic's recent Max plan structure hints at the second path. OpenAI has been quieter about it.

Looking ahead

The repos themselves will get taken down, renamed, or forked into the long tail. That part doesn't matter. What matters is that GitHub Trending is now surfacing AI subscription arbitrage as its own category, which means the price-value gap is wide enough that unrelated developers are independently arriving at the same solution. That's a market signal, and vendors read GitHub Trending too. The interesting question for the next two quarters isn't whether the grey market grows — it will — but which vendor blinks first on pricing, and whether the skill-packaging approach eats enough per-call volume to make the sub2api approach unnecessary. If you're building AI infrastructure right now, plan your unit economics for both outcomes.

GitHub 693 pts 68 comments

LinXiaoTao/FuckClaude: New trending repository

→ read on GitHub
GitHub 544 pts 39 comments

dzhng/skills: New trending repository

→ read on GitHub
GitHub 248 pts 106 comments

akihitohyh/chatgpt-register-sub2api: New trending repository

→ read on GitHub

// share this

// get daily digest

Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.