Anthropic confirmed that three engineering missteps in March and April 2026 caused significant performance degradation in its Claude AI, following weeks of user complaints about "AI shrinkflation." The issues included reduced reasoning effort, a caching bug that made the model "forgetful," and a verbosity limit that hurt code quality. Anthropic has since reverted the changes and reset user limits, promising more transparency.
Claude Opus 4.7 release
- ▪Anthropic confirmed three distinct changes in March and April 2026 caused performance degradation for users of Claude Code, Claude Agent SDK, and Claude Cowork.
- ▪Claude Opus 4.7 was released with improvements to agentic coding, a new tokenizer, and vision updates.
- ▪Anthropic explained the degradation was caused by changes to the model's 'harness' and 'operating instructions,' not the core model weights.
- ▪Users on platforms like GitHub, X, and Reddit reported a perceived degradation in Claude's performance, describing it as "AI shrinkflation."
Agentic persistence improvements
- ▪Opus 4.7 substantially fixes a 'persistence deficit' from Opus 4.6, reducing task abandonment rates by roughly 60%.
- ▪The caching bug, which affected Sonnet 4.6 and Opus 4.6, was fixed on April 10, 2026.
- ▪A bug introduced on March 26 caused Claude's cached session data to be cleared with every turn, making the model "forgetful and repetitive."
Coding benchmark performance gains
- ▪Claude Opus 4.7 showed improved scores on coding benchmarks, with SWE-Bench Verified rising to 78.9% from 71.2% and HumanEval to 91.7% from 88.3%.
- ▪On March 4, 2026, Anthropic changed Claude Code's default reasoning effort from "high" to "medium" to reduce latency, a change it later called "the wrong tradeoff."
- ▪Cybersecurity firm TrustedSec reported that Claude's code quality became "unusably bad," dropping by over 47.3% after the Opus 4.6 release.
New tokenizer cost implications
- ▪As compensation for performance issues and token waste, Anthropic reset usage limits for all subscribers on April 23, 2026.
- ▪Opus 4.7's new tokenizer increases token counts for English text by 12-18%, effectively raising costs for English workloads.
- ▪On April 16, 2026, a system prompt change to reduce verbosity caused a 3% performance drop in coding quality before being reverted on April 20.
Web research quality regression
- ▪An analysis by coding security company Veracode found that Claude Opus 4.7 included a vulnerability in 52% of coding tasks.
- ▪While improving in coding, Opus 4.7 is weaker than Opus 4.6 in web research tasks, with decreased accuracy in source synthesis and contradiction detection.
Model comparison landscape
- ▪Veracode's analysis found OpenAI's models performed better than Claude's, introducing vulnerabilities in only around 30% of tasks.
- ▪In response to the issues, Anthropic is implementing safeguards like more internal testing and a new @ClaudeDevs X account for transparency.
- ▪Anthropic acknowledged that its infrastructure has been stretched by unprecedented demand, a constraint across the AI industry.
Story comments
Loading comments…