Codex CLI on Bedrock can't opt into prompt caching, and one user's logs show 85% of a $1,386 bill was cache writes.

GitHub issue openai/codex#37674 documents that Codex CLI 0.147.0 running GPT-5.6-Sol on Amazon Bedrock cannot pass prompt_cache_options or prompt_cache_breakpoint, so every request performs a cache write instead of a read. August 5-8 production logs: 3,656 requests, 171.94M cache-write tokens, an estimated $1,182.09 of $1,386.46 total, roughly 88K wasted cache-write tokens per request in local tests. GitHub Still open. Check your token telemetry before the next invoice if you're on this path.