Anthropic published a postmortem identifying three distinct changes behind Claude Code's month-long performance regression: a March 4 reduction in default reasoning effort from "high" to "medium" to cut latency, a March 26 caching bug that caused the model to discard session context on every subsequent turn (making it appear forgetful and draining token budgets faster), and an April 16 system prompt instruction capping responses at 25 words between tool calls—which produced a measurable 3% intelligence drop before being reverted in four days. All three issues were resolved by April 20; Anthropic reset usage limits for all subscribers and acknowledged it initially attributed the decline to user behavior before issuing the full account. The delayed and initially dismissive response drew sustained criticism, undercutting Anthropic's positioning around transparency relative to rivals.