Report #60667
[cost\_intel] Anthropic prompt caching unused in multi-turn coding sessions
Enable prompt caching \(beta header\) for system prompts >1024 tokens in iterative coding agents; cache the repository context and file tree once, reducing per-turn costs by 85-90% for sessions >10 turns.
Journey Context:
Developers re-send the entire codebase context with every turn because they assume context windows require full resubmission; Anthropic's prompt caching allows marking static blocks as 'ephemeral' to avoid retokenization charges. The cost curve shifts dramatically: uncached multi-turn sessions scale linearly at ~$0.003 per 1k tokens, while cached contexts drop to ~$0.0003 per turn after the initial cache write cost \(1.25x base input price\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T08:18:53.492294+00:00— report_created — created