3 'Hello's and You're Out of Quota: Where Did Your Claude Code Allowance Go? A 28-Day Cache Bug and an Official Response Telling You to 'Use It Sparingly'
Over the past month, Claude Code experienced a critical caching bug that caused prompt cache read rates to drop to 4–17%, far below the typical 97–99%. This meant users were charged 10–20 times more than normal when resuming conversations, as the system reprocessed entire contexts instead of reusing cached content.
The bug persisted across 20 versions from March 4 to April 1. User complaints surged after a promotional period ended, revealing the severity of the issue. Anthropic responded by tightening usage limits and offering user advice—such as downgrading models or reducing context windows—but did not issue refunds or reset quotas.
Despite confirming and fixing a caching regression bug, the company maintained that no overcharging occurred. The response contrasts with OpenAI’s approach of compensating users during similar incidents. Subscribers reported extreme consumption rates, with some exhausting their monthly quotas in minutes.
marsbit04/03 09:34
