Claude Code accounts for 4% of all public GitHub commits. 135,000 per day. By any product metric, it’s winning.

OpenCode went from 40,000 to over 125,000 GitHub stars in three months. 2.5 million monthly active developers. And the event that triggered that growth was Anthropic blocking third-party access to Claude on January 9, 2026.

Anthropic built the best coding tool on the market and then created its biggest competitor’s best marketing campaign.

The switch

At 2:20 AM UTC on January 9, Anthropic flipped a switch. Third-party tools using Claude through subscription OAuth stopped working. No warning. No migration path.

Tools like OpenCode had been sending headers that made Anthropic’s servers think requests came from the official Claude Code client. From Anthropic’s perspective: spoofing. From users’ perspective: using a model they paid $200/month for through an interface they preferred.

The open-source community found workarounds within days. Anthropic blocked those too.

On February 19, the Terms of Service were updated: OAuth tokens from Free, Pro, and Max plans may not be used with third-party tools. The same day, OpenCode pushed a commit removing all Claude OAuth code. The commit message cited “Anthropic legal requests.”

By March, Anthropic took direct legal action against OpenCode.

The walled garden was sealed.

The gym membership problem

The economic argument is legitimate. Let’s be fair.

Claude subscriptions are subsidized. $200/month gives you tokens at a fraction of API pricing. That model works because most subscribers don’t max out — the gym membership model, profitable because most members don’t show up every day.

Third-party tools broke that assumption. Autonomous coding loops running overnight. Agents burning tokens while users slept. The casual gym members were suddenly showing up at 3 AM using every machine.

Users running Claude through the API report spending $6 on a single small feature. An intensive month could hit $1,000. At $200 flat, Anthropic was losing money on every power user.

So the block makes financial sense. But financial sense and strategic sense aren’t the same thing.

Two CEOs, same data, opposite moves

Two open-source projects were disrupting the market:

OpenClaw — originally named Clawdbot (yes, after Claude) — was an AI agent framework built by Austrian developer Peter Steinberger as a weekend project. By mid-February 2026: over 200,000 GitHub stars and climbing vertically toward becoming the most-starred software project on GitHub.

OpenCode — an open-source terminal coding agent supporting 75+ model providers. Key value proposition: no vendor lock-in.

Both projects drove massive traffic to Claude. Both were growing faster than Anthropic’s own tools.

Dario Amodei sent trademark complaints to OpenClaw, forcing two name changes. Sent legal notices to OpenCode. Blocked OAuth. Updated ToS. Filed legal action.

Sam Altman offered to hire Peter Steinberger. Brought him to OpenAI to lead next-generation personal agents. Opened subscriptions to third-party tools. Partnered with OpenCode.

Translation: Anthropic’s lawyers sent the cease and desist. OpenAI sent the offer letter.

Results: OpenAI absorbed the creator of the fastest-growing open-source AI project in history and his entire community. Anthropic got called “customer hostile” by the creator of Ruby on Rails and generated OpenCode’s biggest growth event — 18,000 new stars in two weeks.

One CEO saw a cost problem. The other saw a distribution opportunity. Same data.

The accounts that mattered most

The clearest signal was the accounts being banned after January 9.

$200/month Max subscribers. Power users. Content creators who made videos recommending Claude. Developers who brought Claude into their teams. The highest lifetime value customers with the highest influence in the ecosystem.

Banned for logging into Claude through a different interface.

YouTube creator Gentleman Programming — someone who actively promotes Claude to thousands of Spanish-speaking developers — made a video titled “SE PUDRIÓ TODO” (Everything went to hell). 2,300 views in two hours. That’s 2,300 developers who now associate Anthropic with restriction instead of innovation.

The cost of acquiring a developer’s trust takes years. The cost of losing it takes one YouTube thumbnail.

The product isn’t the problem

Claude Code has the highest accuracy on SWE-bench Pro. Its checkpoint system, subagent architecture, and explore agent are best-in-class. Users didn’t leave because Claude Code was bad.

They used third-party tools because those tools offered something Claude Code didn’t: choice. Model switching. Custom workflows. The ability to use Claude — the model they preferred — through an interface that better fit how they worked.

That’s not disloyalty. That’s the market telling you something. And the correct response to the market telling you something is not to sue the market.

The Pentagon parallel

The same dynamic played out at a larger scale in February.

Anthropic refused to remove safety guardrails from Claude for Pentagon use in autonomous lethal warfare and mass surveillance. Trump called them a “radical left, woke company” and ordered agencies to immediately cease using Anthropic technology. The Pentagon designated them a “supply-chain risk to national security.”

Hours later, OpenAI struck a deal with the Pentagon.

With the military, Anthropic took a principled position on ethics. With the developer ecosystem, they took a principled position on Terms of Service. One is genuinely about safety. The other is about margins. But the market charges the same price for both.

What should have happened

The alternative wasn’t “let everyone arbitrage tokens forever.” But between “do nothing” and “send lawyers,” there was a middle:

A partnership program. Blessed third-party tools get subscription-tier access in exchange for branding and revenue sharing. OpenCode displays “Powered by Claude.” Anthropic gets distribution it couldn’t build alone.

A developer tier. $300/month that explicitly allows third-party usage with rate limits. Capture the willingness to pay instead of banning it.

An acquisition. OpenCode was at 40,000 stars when the drama started. The cost in January would have been a rounding error. Instead, Anthropic created a competitor now at 125,000+ stars.

OpenAI demonstrated all three: hired the OpenClaw creator, partnered with OpenCode, opened subscriptions. Not because they’re more generous — because they understood that when models converge, distribution wins.

The paradox

Anthropic has the best model. The best coding tool. The most principled safety position. And they’re building walls in a market moving toward doors.

The paradox isn’t that they’re wrong about the product. They’re right — Claude is remarkable. The paradox is that being right about the product doesn’t protect you from being wrong about the market.

The control that preserves margins today erodes the ecosystem that generates revenue tomorrow. And by the time the impact is measurable, developers will have built their workflows around other models — not because those models are better, but because those models let them work the way they want.

Every wall you build keeps customers out as much as it keeps them in.

On January 12, 2026, OpenCode hit a peak of 2,087 new GitHub stars in a single day. The market was voting. Anthropic responded with lawyers. The market kept voting.