Anthropic in April 2026: Record Revenue Crushes Infrastructure, Mythos Terrifies Bankers

The article maps the most turbulent period in Anthropic's history in spring 2026, when a surge of users following the #QuitGPT movement and a dispute with the Pentagon brought record revenue (over $30 billion annualized) and a $30 billion investment at a $380 billion valuation, but also exposed critical capacity limits — dozens of outages, tightened rate limits, and a demonstrable decline in Claude Code quality documented by AMD's analysis. The author also examines the legal dispute with the U.S. Department of Defense, the secret Claude Mythos model, and the interplay of five factors (inadequate infrastructure, quiet cost optimization, reasoning redaction, error-prone iterations, and lagging QA) behind the service failures. In closing, he assesses whether the company can stabilize its infrastructure, restore developer trust, and achieve a clear legal outcome before its planned IPO in autumn 2026.

Anthropic is going through the most turbulent period in its history. The explosion in user numbers following the #QuitGPT movement and the dispute with the Pentagon brought record revenue, but at the same time exposed serious capacity limits in its infrastructure. Claude's services have logged dozens of outages over the past 90 days, the quality of Claude Code has demonstrably declined, and the company faces an unprecedented legal dispute with the U.S. Department of Defense. Despite all this, in February 2026 Anthropic secured a $30 billion investment at a $380 billion valuation and is preparing to go public.

Service outages reach critical frequency

On April 6–7, 2026 alone, Claude logged two large-scale outages that generated thousands of reports on Downdetector. In the first ten days of April there were at least eight separate incidents — ranging from elevated API error rates through authentication failures to complete unavailability.

The status.claude.com page documents incidents on an almost daily basis. Over the past 90 days, Claude.ai has achieved only 98.73% availability, while the API is doing slightly better (98.95%), suggesting that Anthropic is prioritizing enterprise customers. Over the past month, the Statusfield portal recorded 39 incidents and availability of just 89.4%.

The cause is a fundamental imbalance between demand and capacity. After the #QuitGPT movement, Claude was receiving hundreds of thousands of new sign-ups every day, and the app reached the number one spot in the U.S. App Store.

Anthropic responded by tightening limits during peak hours (weekdays 5:00–11:00 a.m. Pacific time), which it announced on March 26 not on its blog but via a tweet from a technical staff member. Roughly 7% of users hit restrictions that previously did not exist. Subscribers to the Max plan, which costs $100–200 per month, report exhausting their limit after just one hour of work. Anthropic admitted that Claude Code quotas were being used up too quickly.

A proven decline in quality: the AMD case

The most serious case was documented by Stella Laurenzo, director of the AI group at AMD, in GitHub issue #42796 dated April 2, 2026. Her analysis covers 6,852 sessions, 234,760 tool calls, and 17,871 thinking blocks.

The median depth of the model's "thinking" dropped by 67% — from roughly 2,200 characters in January to just 720 characters at the end of February. The ratio of file reads to edits fell from 6.6 to 2.0 — the model began modifying code without first reading it properly. User frustration and the number of forced interruptions also rose noticeably.

Anthropic responded through engineer Boris Cherny, who explained that the redact-thinking setting merely hides the reasoning from the interface and does not reduce the model's actual thinking depth. The company admitted to introducing "adaptive thinking" as the default setting from the release of Opus 4.6 on February 5 — the model dynamically determines for itself how much time to devote to reasoning.

The financial impact on AMD was drastic: the team's bill jumped from $345 to $42,121 per month due to repeated fixes of faulty outputs. A key finding was that in some cases the model allocated a zero thinking budget, which led to fabrications.

In addition, on March 3 the default effort level was lowered from "high" to "medium," reportedly as a compromise between intelligence and latency. The recommended fix: set the environment variable CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 or use the /effort max command. Anthropic subsequently restored the default effort to "high" for the API, Bedrock/Vertex, and enterprise users, but closed the issue on April 8 without a full explanation.

Claude Mythos Preview: a new, most capable model for the chosen few only

On April 7, 2026, Anthropic announced Claude Mythos Preview — described as the most capable model they have ever created. It introduces a new tier, Capybara, sitting above the existing Opus.

The price for participants in the Project Glasswing program is $25/$125 per million input/output tokens, that is five times more than Opus 4.6. The model is available exclusively by invitation — about 50 organizations have access. Anthropic invested $100 million in credits and at least $2.5 million in security organizations under the Linux Foundation (Alpha-Omega and OpenSSF).

Mythos demonstrated extraordinary capabilities in cybersecurity: it autonomously identified thousands of zero-day vulnerabilities, including bugs as old as 27 years. Anthropic declared that the model will not be made publicly available — it is considered too dangerous for unrestricted deployment.

The reaction was unprecedented. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened an emergency meeting with the heads of the largest U.S. banks. Bloomberg reported on it as an urgent warning. Present at the meeting at the Treasury Department's headquarters were representatives of Citigroup (Jane Fraser), Morgan Stanley (Ted Pick), Goldman Sachs (David Solomon), Bank of America (Brian Moynihan), and Wells Fargo (Charlie Scharf). JPMorgan chief Jamie Dimon was invited but could not attend — JPMorgan is at the same time a partner in Project Glasswing.

The relationship between the Mythos brand and a possible Claude 5 remains unclear. Community speculation estimates a Claude 5 release in the second or third quarter of 2026, but these are only guesses. The most recent publicly available models remain Opus 4.6 (February 5) and Sonnet 4.6 (February 17, 2026).

The dispute with the Pentagon: the legal landscape splits

The conflict between Anthropic and the U.S. Department of Defense is one of the most closely watched technology-political disputes of 2026.

In February 2026, Defense Secretary Pete Hegseth gave CEO Dario Amodei an ultimatum — the Pentagon demanded the use of Claude for all lawful purposes without restriction. Amodei publicly refused, and Anthropic insisted on two red lines: no fully autonomous weapons and no mass domestic surveillance.

Trump ordered all federal agencies to immediately stop using Anthropic technologies, and Hegseth labeled the company a "supply chain risk" — a designation historically reserved for foreign adversaries. OpenAI secured a Pentagon contract for a classified network that same day. ChatGPT uninstalls rose by 295%, and the #QuitGPT movement brought millions of new users to Claude.

The legal situation is divided. On March 26, Judge Rita F. Lin issued a preliminary injunction blocking the government's designation of Anthropic as a security risk and the ban on the use of Claude by government agencies. By contrast, on April 8 the D.C. Circuit Court of Appeals rejected Anthropic's request to stay the supply chain risk designation specifically for Department of Defense purposes. The result: Anthropic cannot work with the DoD but can continue with all other government agencies. Oral arguments in the D.C. Circuit are scheduled for May 19, 2026.

The Pentagon now has no access to the Mythos model, which is considered strategically crucial for cyber defense — commentators describe the situation as a "strategic own goal" by the administration.

Financial position: record revenue, record spending

In February 2026, the company closed a Series G investment round worth $30 billion — at a $380 billion valuation, more than double the $183 billion from September 2025. The round was led by GIC and Coatue, with D.E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX as co-leads. Other investors included Sequoia Capital, BlackRock, Blackstone, and dozens more. In total, the company has raised roughly $64 billion since its founding. It was the second-largest venture investment in history, behind last year's OpenAI round ($40 billion).

Revenue is growing at an unprecedented pace. From $1 billion in annualized revenue in December 2024, the company climbed to $9 billion by the end of 2025, $14 billion in February 2026, $19 billion in March, and an estimated $30 billion in April 2026 — thereby surpassing OpenAI and its roughly $25 billion for the first time. Claude Code alone generates over $2.5 billion a year; enterprise subscriptions have quadrupled since January.

In addition, in late March the complete source code of Claude Code leaked (512,000 lines of TypeScript across 1,900 files) due to a missing .npmignore in npm package version 2.1.88. Security researcher Chaofan Shou uncovered a 59.8 MB source map, which was mirrored on GitHub within hours. Anthropic described the incident as a "packaging error caused by human error."

Key partnerships with cloud providers continue: Amazon has invested $8 billion in total, while Microsoft and NVIDIA plan to invest up to $15 billion. Anthropic announced an expansion of its partnership with Google and Broadcom on custom chips for inference. New compute capacity is expected to reach 3.5 gigawatts from 2027. More than 1,000 companies spend over a million dollars a year on Claude.

The company is actively preparing to go public — according to sources at Let's Data Science and others, it is targeting October 2026 on the Nasdaq. The timing and the valuation may be affected by the situation surrounding the Mythos model and its security implications. On the other hand, planned spending for 2026 amounts to roughly $19 billion, and gross margins of around 40% are below the level needed for long-term sustainability. OpenAI projects a loss of $14 billion for 2026; Anthropic is targeting positive free cash flow by 2027.

Why the services are failing: the interplay of five factors

The decline in service quality and speed is not caused by a single factor, but by a confluence of mutually reinforcing ones.

First, infrastructure capacity is not keeping up with demand. The millions of new users from the #QuitGPT migration flooded the systems faster than Anthropic could scale GPU/TPU capacity, leading to extreme latency — sometimes 15 seconds or more to the first token.

Second, cost optimizations quietly reduced quality. The shift to adaptive thinking and the lowering of the default effort were probably motivated by the need to save compute capacity, but had a devastating impact on complex tasks — as the AMD case showed.

Third, the redaction of thinking blocks made it impossible for users to diagnose problems, which created a crisis of trust among developers and demanding users.

Fourth, rapid iterations introduced bugs. A prompt cache bug from March 23 increased costs by up to tenfold to twentyfold. A npm packaging error led to the leak of the Claude Code source code on March 31.

Fifth, quality assurance did not keep pace with the speed of development. In March, Anthropic shipped 14 product updates alongside 5 outages — a ratio that points to a systemic problem in the development process.

The fundamental capacity problem remains unresolved and will likely take months before the new compute resources come fully online.

Conclusion

In April 2026, Anthropic finds itself in a position comparable to Twitter's "fail whale" era — enormous success crushing the infrastructure. The company is a technology leader with the most capable models and record revenue, but it pays for this with service instability and an erosion of trust within the developer community.

The key question will be whether, before its planned IPO in the fall of 2026, Anthropic can resolve three things: stabilize the infrastructure, restore developer trust, and achieve a clear legal outcome in its dispute with the government. Mythos suggests that at the level of pure AI capabilities, Anthropic is at the cutting edge — the question that remains is whether it can deliver that cutting edge reliably to its users.

The concept, structure, and editorial line of the article are the work of the author, who prepared the content outline, set out the key theses, and directed the entire creative process. Generative AI (Claude, Anthropic) was used as a tool for research, fact-checking, and fleshing out the author's draft.

The author edited the outputs throughout, verified the key findings, and approved the final wording. No part of the text was published without human oversight. All factual data were verified against the publicly available sources cited in the text.

The procedure complies with the requirements of Art. 50 of EU Regulation 2024/1689 (the AI Act) on transparency of AI-generated content. #poweredByAI