Anthropic details Claude Code regressions

This is a useful postmortem because it admits the uncomfortable truth: you can make a frontier coding agent feel dramatically worse without changing the base model at all. For developers, that is a reminder that harness, prompting, caching, and defaults are now part of the product's real capability surface.

• –The biggest takeaway is that perceived “model nerfs” can come from UI and orchestration decisions, not model weights, which makes observability and rollback discipline much more important for AI tools.
• –Anthropic’s stale-session bug is especially notable because it degraded both quality and quota efficiency at once, making Claude Code look forgetful while also burning usage limits faster.
• –The prompt change is a warning shot for anyone shipping agent products: aggressive brevity constraints can save tokens but still tank coding performance if they suppress planning and tool-use context.
• –Anthropic’s writeup also suggests its public eval coverage lagged real-world usage; the company is now promising broader per-model prompt evals, longer soak periods, and more staff on exact public builds.
• –Community reaction appears split between frustration over how long this took to diagnose and respect for Anthropic publishing a specific, technical postmortem instead of hand-waving away the complaints.