Source · docs
OpenAI Ultrafast mode guide (GPT-6.1 Sol support)
Summary
Ultrafast is OpenAI's fastest API service tier, meant for applications where lower latency between output tokens is worth a higher price. OpenAI's API changelog of 2026-10-08 added GPT-6.1 Sol to it; GPT-6 Astra already had it (since 2026-09-29). The guide is undated; the publication date here is the changelog entry.
Key points
- Supported models (as of 2026-10-09): GPT-6 Astra and GPT-6.1 Sol in full, GPT-5.6 Sol as a preview.
- Enable per request in the Responses API with
service_tier: "ultrafast"and modelgpt-6-astraorgpt-6.1-sol. - Separate rate limits from the Standard and Fast tiers; check them before scaling.
- Data residency: GPT-6.1 Sol supports global processing plus US and EU residency; GPT-6 Astra only global and US.
- OpenAI "strongly recommend[s] WebSockets, especially for agentic applications" to cut network overhead.
- Prices are on a separate Ultrafast pricing table (input, cached input, cache write, output); the guide gives no speed figures.
Why it matters
Output speed is becoming a priced option on the same model, like Anthropic's fast mode for Opus. For interactive agents, latency can now be bought rather than obtained only by switching to a smaller model.