/v1/messages. Codex speaks OpenAI protocols only, so Claude there falls back to chat/completions compatibility mode — no cache_control, no cache_read_input_tokens, and the full context is billed at list price on every turn. It is the easiest trap to fall into when your Claude Code subscription runs out of quota and you try to continue inside Codex: you save the account switching, but the bill goes up.
If Codex really must orchestrate Claude, have it generate and run code or an MCP tool that calls /v1/messages; caching works on the native format. The Claude model page and prompt-caching guide now explain the details.
See Claude API Basics and the Claude Prompt Caching Guide.
← Back to Live Updates · 📚 Monthly archive