Anthropic downgraded cache TTL on March 6th

Article Summary

Anthropic silently downgraded the prompt cache TTL in Claude Code from 1 hour back to 5 minutes around March 6, 2026, without any changelog or announcement. This caused a dramatic cost increase for API users — cache writes are 12.5x more expensive than cache reads — with one user calculating $949-$1,582 in overpayments across ~120K API calls.

Discussion

• Users report Claude Code became “virtually unusable” in mid-March as session quotas started expiring within the first hour due to expensive cache re-creation
• Several commenters note a broader pattern of silent quality/capability degradation at Anthropic, with engineers increasingly discussing Claude unfavorably
• Speculation the TTL downgrade may be linked to Anthropic’s compute constraints, with peak-hour throttling as a downstream symptom
• A wave of users switching to competitors, with one noting “Codex is absolutely fantastic right now” after switching from Claude

Discuss on HN