Anthropic lowered the cache TTL on March 6

After analyzing 119,866 JSONL session logs from ~/.claude/projects/, evidence is presented that Claude Code's default prompt cache TTL quietly changed from 1 hour (1h) to 5 minutes (5m). The timing of the change is estimated to be around March 6–8, 2026.

The same pattern was observed on two different devices and accounts.

• January 11 – January 31: only 5m seen
• February 1 – March 5: only 1h seen consistently
• March 6–7: 5m starts appearing again
• After March 8: 5m becomes dominant

The author claims this change appears to be a server-side configuration change. The basis for this is that the client version and usage patterns did not change, and the TTL tier is controlled on Anthropic's side.

The cost impact was also calculated to be significant.

• Based on Sonnet, total actual cost of $5,561.17 vs. $4,612.09 (assuming 1h was maintained)
• $949.08 in excess spending, 17.1% wasted
• Based on Opus, the same ratio results in $1,581.80 in excess spending

The core logic is as follows.

• During the 1h TTL period, there were almost no 5m tokens, and February had virtually no waste
• Starting in March, as 5m tokens surged, cache creation costs and quota consumption increased together
• For Claude Code's usage pattern with many long sessions, a 5-minute TTL is seen as excessively disadvantageous

The author requests the following from Anthropic:

• Confirm whether the 1h default in February 2026 and the regression to 5m in March actually occurred
• Disclose Claude Code's intended default TTL
• Restore the 1h TTL or provide a user-configurable option
• Disclose how quota is calculated for cache_read tokens