Codex bug: GPT-6.2 cache hit rate drops from 97.4% to 46.4%, system prompts billed at full price
steipete · x · 2026-11-20
A billing-grade caching bug was found in OpenAI's Responses API: on GPT-6.2 Astra and Sol, content in the top-level instructions field is never actually cached — 0 cached tokens every turn — while OpenAI's own diagnostics incorrectly report a cache hit (Codex issue #62312).
The impact is large: moving from GPT-6.1 Sol to 6.2 Sol dropped real cache hit rates from 97.4% to 46.4%, meaning 64K–72K system tokens get billed at full price on every step.
Workaround: OpenCode puts the system prompt inside a leading developer message in input. steipete has already landed a fix in OpenClaw; most harnesses aren't updated yet.
More from coding & agent
- Vercel's new UI SDK draws praise from shadcn for building draggable calendar, date picker — shadcn · 2026-11-20
- Dev shows AI meal-planning tool that matches macros using whole grocery units — lucasmeijer · 2026-11-20
- Agents run animal content accounts to 300M-impression marketing playbook — Saul_Loveman · 2026-11-20
- Dev ships agent-built recipe site with copyable agent instructions — lucasmeijer · 2026-11-20
- Cost-vs-quality chart puts Gemini 4 Neon on the CursorBench Pareto frontier — zacharynado · 2026-11-20
- 40 lines of zero-dependency Node: AES-256-GCM encrypt env vars before your PaaS leaks them — DavidWells · 2026-11-20