Overview
- OpenAI announced the Ultrafast preview on August 13 and is offering it first through the API to a select set of enterprise testers on a waitlist.
- The company says Ultrafast runs GPT‑5.6 Sol up to 14 times faster than standard processing and can produce as many as 750 output tokens per second.
- Cerebras supplies the wafer‑scale hardware that OpenAI credits for the lower latency, and OpenAI’s own developers are already using the tier for incident response and faster research loops.
- Early customers including Jane Street, Podium, Basis and Rogo are trialing Ultrafast for live voice, support, commerce, coding agents and financial research.
- Industry observers warn access is capacity‑constrained, that vendor speed figures vary by workload, and that businesses should run representative tests, document model choices and plan staged adoption as vendors split AI into fast/expensive and slow/cheap lanes.