GPT-6.2 breaks Responses API prompt caching: hit rate falls from 97.4% to 46.4%
lxfater · x · 2026-11-22
Developers report that GPT-6.2 changed Responses API prompt caching behavior: fixed rules placed in top-level instructions are now moved to the end of the conversation, so they never hit the cache and get re-billed at full price every turn.
- Cached tokens previously cost 5% of full price ($2 → $0.1 per M tokens, a 20x difference)
- Data from the Codex GitHub repo: with the same tooling and rules, 6.1 Sol ran 13k+ turns at 97.4% cache hit rate; 6.2 Sol dropped to 46.4% over 122 turns
- Third-party tools logged in via ChatGPT subscription also burn quota faster
Official workaround: stop putting fixed rules in top-level instructions; put them as the first item of the input list with the developer role. OpenCode verified that after the move, cache hit rates for both Sol and Astra return above 97% from the second turn.
How to check: inspect cachedtokens in your logs — if the cached share suddenly collapsed, check where your rules sit, or update your tooling.
More from coding & agent
- t3code runs 8 sub-agents on the same brief in parallel with one sentence — jasonkneen · 2026-11-22
- Dev builds auto-generated collaborative canvas to iterate on creative work with AI agents — max__drake · 2026-11-22
- Rustingstrata + DSH balloons prompts to 195k tokens while llama.cpp works fine — Mordimer86 · 2026-11-22
- Open-source Laya: single-pass System 1 decision engine that keeps learning from LLMs — solyarisoftware · 2026-11-22
- CTO merges branch 243 commits ahead of main, flexing second GitHub YOLO badge — mayanicks0x · 2026-11-22
- Dev uses Codex to relight Elden Ring into a next-gen look — runs on a 2016 GTX 1070 — banteg · 2026-11-22