Overview
- Anthropic said people are hitting Claude Code usage limits far faster than expected and called an investigation its top priority.
- Developers report trivial interactions burning through caps, including greetings counted as 2% of a session and Max allowances depleted within an hour.
- Recent policy shifts changed how limits apply, with a two-week off-peak doubling ending March 28 and new peak-hour reductions expected to affect about 7% of users.
- A reverse-engineering post claimed prompt-cache bugs inflate costs by 10–20x, which Anthropic says it is examining as unconfirmed while some users report relief after downgrading the client.
- Prompt caching can cut repeat costs but the default cache lasts five minutes and longer caching carries higher token prices, and unclear per-plan quotas make budgeting and automated retries risky.