ADR 0166 — A spoken turn asks OpenAI for its Fast service tier

This decision record is in English only.

Status: Accepted · Date: 2026-10-04

Context

The wait between Allen’s words and the first spoken word is dominated by the first token of the model’s first request. On gpt-6-luna through the Responses API, 16 paid calls (2026-10-04) with service_tier: "priority" came back labelled service_tier: "fast" (OpenAI renamed Priority to Fast) and cut the first-token median from 1.04 s to 0.68 s and the maximum from 1.86 s to 0.85 s.
Fast is billed at a premium. The 2x figure (gpt-6-luna input 0.20, cache read 0.02, output 1.00 USD per 1M tokens) is second-hand and not checked against OpenAI’s pricing page; the cache-write rate is an assumption.
Only a spoken turn is waited on. Daily reports, mail and job-mail jobs, the day summary, core memory, compaction, projects, vision, dictation, the prefix warm and every other background request have no listener, and a typed turn has a screen to wait on.
The field is OpenAI’s own. Sent to OpenRouter (Jev), x.ai or Anthropic it is at best ignored and at worst a rejected request.
cost.recorded prices a request from its model row. The tier the provider says it served is what it bills, and can differ from the tier asked for.

Send realtime.response.voice_service_tier as the service_tier of every model request decide() makes for a turn Allen spoke.

Its limits:

Off by default. An empty or absent key sends nothing, and the request body is then unchanged.
Voice turns only. The composition root sets it on the turn’s DecideContext only for a spoken channel (not typed, not GPT-Live). It covers the spoken stream and its tool-loop follow-ups, the routine stream, the batch loop, the answer after the tool budget and the spoken-form rewrite, which belongs to the voice answer. Nothing else reads the key.
OpenAI’s host only. The client drops the field unless the provider is OpenAI and the base URL is api.openai.com, whatever the caller passes.
Per request. The tier is an argument of one request, captured into its body at prepare time; it is never stored on the client or a preset.
Billed as served. The cost recorder prices a request at the tier the response reports, else the tier asked for, from a <model>:<tier> row in data/pricing.json (priority is fast); a tier with no row bills at the model’s standard row. cost.recorded carries the tier as the optional service_tier.

Set the tier on the preset. The preset serves every kind on that model, so daily reports and mail jobs would pay the premium for latency nobody waits for.
Send it for every request of a model. Same cost, and it cannot be kept off OpenRouter or x.ai by configuration alone.
Keep the standard price for fast requests. cost.recorded would run at half the real rate on exactly the turns the switch changes, and the cost guard and every report built on it would under-count.

A turn on Fast costs about twice as much per token until the price row is checked against OpenAI’s page; the row’s source says it is unverified.
A model with no <model>:fast row is billed at its standard rate on Fast, an undercount, until a row is added.
Fast is an OpenAI capacity tier: whether it keeps its latency under load, and its rate limits, are not measured here.