Anthropic Explains Claude Code Degradation and Resets User Rate Limits

Anthropic's engineering team has published an in-depth postmortem dissecting the Claude Code issues developers spent the past month reporting. An internal investigation confirmed that the collective sense of model degradation was not subjective. It was triggered by the overlap of three isolated changes deployed to the codebase between early March and mid-April. All issues affected the Claude Code interface, Claude Agent SDK, and Claude Cowork, while leaving the underlying raw API unaffected. By April 20, developers fully patched the bugs in build v2.1.116 and reset usage counters for all subscribers as compensation.

The latency vs. reasoning depth trade-off trap

The first hit landed on March 4, when Anthropic switched the default reasoning effort for Opus 4.6 from high to medium. Engineers were trying to mitigate latency: under peak load, generation took too long and the terminal appeared frozen. Internal synthetic benchmarks suggested the difference was negligible, but real-world coding suffered, showing dropped accuracy on complex algorithmic tasks. Users rarely tweaked settings manually via the console, leaving them stuck with the downgraded configuration. On April 7, Anthropic reverted the default and pushed Opus 4.7 to the maximum xhigh level.

Cascading memory wipes across long sessions

A far more dangerous caching bug slipped in on March 26. To conserve tokens and speed up resumption after an hour of session idle time, the system was designed to clear old reasoning blocks once via the clear_thinking header. Instead of running a single purge, the flag remained active for the rest of the session. Starting from the second interaction, Claude Code discarded the reasoning chain on every step. If a user issued a command midway through tool execution, the model immediately lost track of why it took previous actions. This triggered repetitive loops, erratic edits, and rapid quota burnout caused by constant cache misses.

Key metrics and incident timeline:

• March 4 — Default reasoning effort dropped to medium to combat UI freezing;
• March 26 — Cache optimization deployed with a critical bug that continuously wiped reasoning context;
• April 10 — Hotfix v2.1.101 released, resolving agent memory loss in long sessions;
• April 16 — Strict system prompt limits introduced (up to 25 words between tool calls and up to 100 words in the final output);
• April 20 — Prompt directive rolled back after detecting a 3% quality regression on comprehensive benchmarks, and v2.1.116 shipped;
• Affected configurations — Claude Sonnet 4.6, Opus 4.6, and Opus 4.7 in client environments.

Repressive system prompts and loss of coherence

The third factor was an attempt to curb the verbosity of the new Opus 4.7. Preparing for an April 16 release, developers injected strict limits into the system prompt: no more than 25 words of explanation between tool calls and no more than 100 words for the final response. Pre-rollout internal tests showed no degradation, but comprehensive retrospective ablations revealed a 3% drop in overall quality across both flagship models. Constrained by the brevity limit, the model stopped thoroughly analyzing architectural decisions and dropped contextual explanations that were critical during multi-step refactoring.

Anthropic's postmortem highlights the fragility of coupling agent tooling, caching, and system instructions. Quiet attempts to trim inference costs by pruning context and capping reasoning inevitably break autonomous agent logic. Developers building their own LLM-based solutions should rethink their monitoring strategies: synthetic benchmarks miss critical regressions whenever test suites fail to account for compounding memory failures in realistic, multi-step workflows.