Anthropic Fixes Claude Code Performance Regression

Anthropic investigated complaints that Claude Code had become noticeably worse and found three separate product-layer changes that together produced a perceptible quality drop. The company confirmed the fixes were deployed in patch v2.1.116 on April 20, 2026 and said the API and inference layer were not impacted. "We never intentionally degrade our models," Anthropic wrote in its post-mortem.

The three root causes were changes to default settings, a session caching bug, and an added system prompt. Each impacted different slices of traffic and model variants, producing inconsistent user experience.

• •On March 4 Anthropic changed Claude Code default reasoning effort from high to medium to reduce latency spikes in high mode; this tradeoff reduced reasoning depth for Sonnet 4.6 and Opus 4.6 and was reverted on April 7.
• •On March 26 a background change intended to clear older "thinking" for idle sessions introduced a bug that cleared that state every turn for the whole session, making the agent forget prior context and repeat itself; fixed on April 10.
• •On April 16 Anthropic added a system prompt instruction to reduce verbosity; combined with other prompt edits this materially harmed coding quality and was reverted on April 20. This affected Sonnet 4.6, Opus 4.6, and Opus 4.7.

These are product-harness issues rather than model-weight changes. That means a stable inference stack can still produce degraded outputs when upstream wrappers, default parameters, session management, or system prompts are altered. Third-party testers and developers reported measurable drops; VentureBeat cited a BridgeMind benchmark slide showing Opus 4.6 accuracy falling from 83.3% to 68.3%, illustrating how surface regressions can be large and sudden when multiple product adjustments interact.

The incident exposes several brittle areas in model deployment pipelines. First, default parameter changes (for example reasoning effort) are a high-leverage knob and should be canaried separately across traffic and evals. Second, session hygiene and caching require stricter end-to-end tests; a change meant to improve latency created a persistent correctness bug. Third, system prompts and verbosity constraints need targeted functional tests that include domain-specific evals such as code synthesis and long-form reasoning. Anthropic reset usage limits and says it will change processes to reduce the chance of similar regressions.

The episode reinforces a recurring theme in production LLM systems: the distinction between model-core stability and product-layer fragility. Competitors and customers watch such incidents closely because they affect trust, SLAs, and benchmark claims. For teams building agents, the takeaway is to instrument and benchmark not only the model API but the full harness, including prompt mutations, default-effort semantics, session state management, and UI-level defaults.

Expect Anthropic to publish follow-up process changes and improved canarying and telemetry. Independent benchmarks and developer reports will be the next signal to confirm restoration of prior quality. Also watch whether other vendors adopt stricter guardrails for default-effort knobs and system-prompt rollouts, since these are common vectors for regressions.