OpenAI's Ultrafast runs GPT-5.6 Sol at 750 tokens a second
The full frontier model, not a distilled one — 5.6× faster on GDP-Val with no quality drop, on Cerebras wafers. Limited preview, no price yet.
Copy markdown
750 tokens a second — and it's the full-size model
OpenAI's new Ultrafast tier serves GPT-5.6 Sol at up to 750 output tokens per second, up to 14× faster than Standard processing. It is not a distilled or smaller variant — same frontier weights, just accelerated.
The speed doesn't cost you accuracy
OpenAI reports a 5.6× end-to-end speedup on the GDP-Val economic benchmark with no quality degradation, and roughly 5× faster than Opus 4.8's Fast mode. One Humanity's Last Exam run finished in 11h11m versus 78h+ at baseline.
Why it's this fast: Cerebras wafers, not GPUs
Ultrafast runs on Cerebras' Wafer-Scale Engine, holding the model's weights in 44GB of on-chip SRAM so tokens pipeline across wafers without leaving the chip. That hardware — not a lighter model — is where the throughput comes from.
The catch: limited preview, no price, no date
Access is gated to a select group of customers through the OpenAI API, expanding as capacity grows. OpenAI published no pricing and no GA date — for now you sign up for updates rather than flip it on in your app.
What 750 tok/s actually unlocks
When a frontier model answers in real time, latency stops gating whole categories: sub-second agent tool-loops, live voice assistants, and incident-response copilots that keep pace with an outage. OpenAI's own targets are voice, fraud detection, and checkout support.
Elsewhere: Cursor makes cloud agents boot 3× faster
Cursor's new Cloud Agents 'builds' pre-clone your repo and install dependencies ahead of time, starting agents up to 3× faster and falling back to the last good build when an environment breaks. It is free and becomes the default for all environments on Aug 17.