Three separate Claude Code changes degraded quality from March 4 to April 20; all resolved in v2.1.116 with usage limits reset for all subscribers.
Key Takeaways
Reasoning effort silently defaulted to medium on March 4; reverted April 7. Opus 4.7 now defaults to xhigh, all other models to high.
March 26 caching bug cleared thinking history every turn after 1hr idle, causing compounding forgetfulness and faster usage limit drain; fixed April 10.
System prompt limiting responses to ≤25 words between tool calls shipped April 16; triggered a 3% eval drop and was reverted April 20.
API layer was unaffected throughout; all three issues were isolated to Claude Code, Agent SDK, and Cowork product surfaces.
Opus 4.7 caught the caching bug in back-test code review; Opus 4.6 missed it. Fix: adding multi-repo context to Code Review.
Hacker News Comment Review
Anthropic’s “corner case” label for 1hr idle sessions was widely contested; many builders described multi-hour planning sessions as their normal workflow, not an edge case.
The deeper trust issue was billing opacity: usage limits draining faster with no explanation while paying for a degraded product landed harder than the technical bugs alone.
Competitive pressure surfaced directly: at least one enterprise account reported evaluating GPT-5.4 after OpenAI offered unlimited tokens through summer, citing consistent reasoning trace quality.
Notable Comments
@everdrive: reported Claude verbalizing its own anti-injection directives mid-session, a distinct symptom not covered by any of the three disclosed bugs.
@bauerd: “Instead of fixing the UI they lowered the default reasoning effort parameter” – frames the March 4 effort change as a product cop-out rather than an engineering fix.
