The Buffet Is Closing: How Anthropic Punished the Developers Who Built Its Reputation
There is a moment in every successful platform’s life when it stops needing the people who made it matter. For Anthropic, that moment appears to be now.
On April 4, 2026, Anthropic cut off subscription OAuth access for all third-party agentic frameworks. No meaningful advance notice. No migration window. One month’s credit, and the door closed. The stated reason was compute abuse. That part was real. But the developers who actually got hurt were not the abusers.
Let’s be clear about who broke the buffet.
OpenClaw was not a developer tool in any serious sense. It was a wish-fulfillment machine for people who want to strike gold without learning to pan for it. Peter Steinberger built a framework that let non-developers pipe Claude into WhatsApp and Telegram and call themselves AI engineers. At its peak, an estimated 135,000 instances were running, hammering a $200 flat-rate subscription with the kind of continuous automated load it was never designed to absorb.1 These were not developers writing production code. They were hobbyists running agents they barely understood against a consumer pricing plan, pocketing the arbitrage, and burning Anthropic’s infrastructure in the process.
Anthropic had every right to shut that down. The math was indefensible. Analysts estimated that heavy agentic users were paying five times less than equivalent API costs, with some gaps reported closer to fifty times. A consumer subscription was quietly cross-subsidizing a class of usage it had never priced for. Pulling the plug on that arrangement was not wrong.
What was wrong was pretending there was only one kind of user on the other side of that decision.
There is a second group. Actual developers. People who write production code, understand rate limits, know why prompt caching matters, and chose tools like OpenCode or other well-built harnesses because those tools are genuinely better for their workflows than Anthropic’s own Claude Code. These developers were not arbitraging anything. They were paying for access, using it responsibly, and building real things. They got the same one-month credit and the same closed door as the wishful thinkers who had never shipped a line of code in their lives.
Anthropic knew the difference. The telemetry exists. The company’s own engineers noted that third-party tools generated unusual traffic patterns, which implies they could see exactly what kinds of usage were happening and at what volumes.2 They chose a blanket ban anyway. That is not a technical limitation. That is a policy choice. And the policy chose enterprise contract revenue over the working developers who had spent the previous two years making Claude the default answer whenever someone asked which model to trust.
The timing makes it worse. Peter Steinberger announced he was joining OpenAI on February 14, 2026. Sam Altman welcomed him publicly. OpenClaw was handed to an open-source foundation with OpenAI’s backing. Anthropic’s terms of service revision landed within weeks. Enforcement followed in April. Steinberger called it plainly: “First they copy some popular features into their closed harness, then they lock out open source.” Anthropic has offered no credible counter-narrative to that sequence.2
This is the story beneath the story. Anthropic is not struggling. It is thriving in ways that would have seemed implausible eighteen months ago. The company grew from $1 billion in annualized revenue in early 2025 to roughly $30 billion by early 2026, a trajectory analysts described as among the fastest in enterprise software history. It closed a $30 billion Series G in February 2026 at a $380 billion post-money valuation. Eight of the Fortune 10 are now Claude customers.
But before accepting those numbers at face value, it is worth asking what the revenue actually is. The $30 billion is real sales, primarily pay-per-token enterprise consumption. That much is clear. What is murkier is how much represents cash versus compute credits flowing in from the same investors writing the equity checks. Amazon, Google, and Microsoft have committed tens of billions to Anthropic, structured partly as cloud compute agreements. Anthropic books those on a gross basis. The SEC, ahead of a likely IPO, may require a different treatment. If a meaningful portion of “revenue” is actually investors paying in server capacity rather than money, the headline number looks considerably different. We find out when the S-1 files. Until then, the number deserves a question mark.3
Even taking revenue at face value, the unit economics do not match the valuation. Gross margins sit at roughly 40 percent today. The $380 billion valuation is priced for 77 percent by 2028. One analyst called that gap “one of the most aggressive margin expansion assumptions ever embedded in a private technology valuation.” Anthropic plans to spend approximately $19 billion on training and inference in 2026 alone, roughly equal to its entire annual revenue. Every dollar earned goes straight back out. Dario Amodei told Fortune, in the same week he closed the largest private funding round in history, that a twelve-month delay in AI progress would make him bankrupt. That is not a man who has solved his cost problem. That is a man whose investors are absorbing the cost problem while the math gets worked out. The developers being pushed off flat-rate subscriptions are collateral in that negotiation. Not partners. Collateral.4
Speaking on the Dwarkesh Podcast, Amodei described data center commitments as potentially “ruinous” if timed wrong. Anthropic has committed roughly $50 billion toward infrastructure. It has signed a $30 billion Azure credit agreement with Microsoft. It is leasing compute from Amazon at a scale that could power a million homes. None of that gets cheaper because working developers prefer a better interface.
The OpenAI/Sora parallel is worth noting, even if imperfect. OpenAI killed Sora in March 2026, dissolving a reported $1 billion arrangement with Disney. Disney learned about the shutdown less than an hour after OpenAI’s morning meeting. The reason was identical logic. Video generation consumed ten to fifteen times the compute of a standard conversation, and the revenue did not justify it.5 When compute becomes the constraint, anything that consumes it without proportional return becomes a liability. The difference is that Sora genuinely was a consumer novelty. The developers Anthropic just cut off were building real things.
The caste system this creates is not subtle. Enterprise customers spending seven figures get custom terms and dedicated support. Developers on pay-as-you-go API get clean access at real prices. Everyone else gets a chat interface and a terms of service document. The Pro subscriber who thought $20 a month meant something. The Max subscriber who thought $200 a month meant something more. The competent developer who chose OpenCode because it was better than Claude Code and was using it exactly as intended. They all got the same letter.
The deepest irony is this. Claude Code holds 54 percent of the enterprise coding market against OpenAI’s 21 percent. That did not happen because enterprise procurement teams woke up one morning and chose Anthropic. It happened because working developers used Claude, trusted it, recommended it, and dragged it into their organizations through sheer advocacy. The community Anthropic just treated as a rounding error was, until very recently, its most effective sales force.
Those developers will find somewhere else to go. Some will pay API rates. Some will move to open-source models with no platform risk at all. Some will move to OpenAI, which now employs the person who built the tool that forced Anthropic’s hand and which has been competing aggressively on exactly this turf. Anthropic’s enterprise base will survive the loss. It is large and sticky and does not particularly care who gets cut.
The question is not whether Anthropic survives losing its developer community. It will. The question is whether the people running it understand what they traded away, and whether they made that trade deliberately or just could not be bothered to make a finer distinction.
Based on the evidence, the second option seems more likely. And that is the part that should actually sting.
Sidebar: The Leak, The Prompt, and The Regulatory Play
On March 31, 2026, Anthropic accidentally shipped the entire source code of Claude Code to the public npm registry. 512,000 lines of TypeScript across 1,906 files, including internal system prompts, unreleased features, and instructions withheld from ordinary users. Anthropic called it a packaging error. The internet did not wait for the explanation.
Engineer and YouTuber Dr. Josh Simmons went through the leaked code and found a conditional in the source: user_type === "ant". If you are an Anthropic employee, you get a different instruction set. If you are not, you get something else. The three instructions reserved for employees are: tell the user when they have a misconception, never claim tests pass when output shows failure, and verify work actually works before claiming done.
Read that again. Verifying that work actually works before claiming done is an optional instruction. One that ordinary users do not receive by default.
The leak also revealed an anti-distillation mechanism that injects fake tool definitions into API responses if Anthropic suspects a competitor is scraping outputs to train a rival model, and an undercover mode that instructs Claude to hide its involvement when working in public repositories. A company that built its brand on safety built a feature explicitly designed to deceive.
The regulatory picture compounds the concern. White House AI czar David Sacks accused Anthropic of “running a sophisticated regulatory capture strategy based on fearmongering.” Anthropic responded by donating $20 million to a pro-regulation PAC and quadrupling its federal lobbying spend to $3.1 million in 2025. The regulatory framework it advocates, tiered rules applying only to the most powerful frontier models, would create compliance costs that a $380 billion company absorbs easily. An open-source alternative does not.
Safety as a moat is not a new strategy. Pharmaceutical companies and defense contractors have used it for decades. The company testifying before Congress about AI safety while withholding basic accuracy instructions from its own users, building deception into its flagship product, and funding the candidates who will write the rules is not a contradiction. It is a business model.
Additional sources: Anthropic Series G announcement · European Business Magazine on OpenAI losses · Data Center Dynamics on Amodei compute comments · PYMNTS on third-party agent ban · DEV Community developer breakdown · apiyi.com on OpenClaw ban timeline · Deep Research Global company analysis
Footnotes
- TNW: Anthropic blocks OpenClaw from Claude subscriptions in cost crackdown
- The Register: Anthropic clarifies ban on third-party tool access to Claude
- Sacra Research: Anthropic reports cloud reseller revenue on a gross basis · Techi.com: SEC may require reclassification of cloud credits ahead of IPO
- Shanaka Anslem Perera: The Growth Miracle and the Six Fractures · Techi.com: Anthropic IPO financial analysis
- ALM Corp: OpenAI Sora Shutdown Explained: Why Disney’s $1 Billion Deal Fell Apart
