Claude users who complained about the AI service producing lower-quality responses over the past month weren’t imagining it.
Anthropic on Thursday published the results of a company investigation that found three distinct changes in March and April made things worse for customers using Claude Code, the Claude Agent SDK, and Claude Cowork. Claude's API, the company says, was not affected.
Users complained bitterly about the quality of Claude's output during March and April, with service availability problems further exacerbating the situation.
Anthropic insists it didn't degrade its models intentionally. Rather, several adjustments went awry, creating the perception of creeping AI incompetency.
First, on March 4, Anthropic adjusted Claude Code's default reasoning effort level from high to medium. Effort level controls how much computational effort the model puts into a particular reasoning task. Anthropic hoped this change would reduce latency resulting from longer cogitation periods.
"This was the wrong tradeoff," the company stated. "We reverted this change on April 7 after users told us they'd prefer to default to higher intelligence and opt into lower effort for simple tasks."
Presumably, lowering the default effort level on Opus 4.6 and Sonnet 4.6 would also have lightened the inference burden—models would "think" less and consume fewer tokens, utilizing limited capacity more judiciously.
The latest Claude Code build, v2.1.118, now defaults to "xhigh" on Sonnet 4.6.
Anthropic’s second misstep was a bug introduced on March 26 when a cache optimization change inadvertently cleared cached session data with every turn of the prompt and response cycle.
Claude caches input tokens for an hour, which benefits users by making sequential API calls faster and cheaper. Company engineers aimed to clear output tokens (thinking sessions) for users idle for an hour, as the cache would no longer be utilized after that much time.
Anthropic’s motive for the change was to reduce the cost of resuming a session by disposing of old, irrelevant thinking traces. Instead, engineers introduced a bug that cleared thinking sessions with each turn, resulting in Claude becoming "forgetful and repetitive." This issue was fixed on April 10 for Sonnet 4.6 and Opus 4.6.
Third, on April 16, Anthropic revised its system prompt, among other measures, in an effort to make Claude models less verbose. The added passage sounded harmless, and following several weeks of internal testing, model quality evaluations had suggested the change was safe. However, upon deployment, it also contributed to the observed performance issues.