Anthropic nerfs models over CapEx

Anthropic said the recent drop in Claude Code quality was caused by three product changes and bugs, not a deliberate model downgrade to cut costs. (anthropic.com) In a postmortem published April 23, Anthropic said the API and inference layer were not affected. It said all three issues were resolved by April 20 in Claude Code version 2.1.116. (anthropic.com) The biggest change came on March 4, when Anthropic lowered Claude Code’s default reasoning effort from high to medium after some Opus 4.6 sessions appeared frozen from long thinking times. Anthropic called that tradeoff wrong and said it reverted the default on April 7. (anthropic.com) A second bug, shipped March 26, kept clearing older context from sessions idle for more than an hour on every later turn instead of once. Anthropic said that made Claude seem forgetful and repetitive until the fix on April 10. (anthropic.com) A third change on April 16 added a prompt instruction to make Claude less verbose. Anthropic said that instruction, combined with other prompt edits, hurt coding quality and was reverted on April 20. (anthropic.com) That matters because the public argument had shifted from “Claude feels worse” to “Anthropic is nerfing models to save money.” Anthropic said it “never intentionally” degrades models and reset usage limits for all subscribers on April 23. (anthropic.com) (venturebeat.com) The underlying technical issue is simple: reasoning effort is the model’s thinking budget. Anthropic’s own Opus 4.7 documentation says higher effort raises intelligence but also increases token spend, latency and cost. (platform.claude.com) So the “nerf” story mixed two separate things: Anthropic acknowledged real quality regressions inside Claude Code, while users and commentators inferred a broader compute-driven downgrade across Claude. The company’s postmortem supports the first claim and rejects the second. (anthropic.com) (venturebeat.com) At the same time, Anthropic is operating under visible capacity pressure. Its public status page logged multiple incidents on April 28 and April 29, including Claude.ai unavailability, authentication problems and elevated errors on Opus 4.7, Sonnet 4.5 and Haiku 4.5. (status.claude.com) Outside reporting has tied those strains to a wider compute crunch across the industry. The Decoder, citing The Wall Street Journal and market data, reported Anthropic API availability at 98.95 percent over 90 days ending April 8 and said GPU prices had risen 48 percent. (the-decoder.com) Anthropic has spent April lining up more hardware. CoreWeave announced a multi-year agreement on April 10 to support deployment of Claude models, with new compute coming online later this year. (investors.coreweave.com) Reuters also reported this month that Anthropic is considering custom chip work and expanding access to Google tensor processing units, or TPUs, as demand rises. That is the backdrop for the Claude debate: not proof of a secret downgrade, but a company fixing real product mistakes while racing to buy more compute. (siliconrepublic.com)