OPENAI · AUG 13 2026API LIMITED PREVIEW
OpenAI’s most intelligent model now streams up to 750 output tokens per second on Cerebras wafer-scale silicon. Until now, real-time speed meant a smaller model. That trade-off is gone.
HEADLINE NUMBERS
Tap a die to flip it. Every card you absorb adds Speed Points — and every figure is straight from the announcement.
ANATOMY
Cerebras WSE‑3. One die, an entire 300 mm wafer. Keep scrolling — the surface opens up.
01 / 06
TO SCALE
WSE‑3 spreads 46,225 mm² of compute across a whole wafer. The largest conventional GPU die — 814 mm² — is the faint rectangle inside it. Drag the loupe.
PLAY THE FIVE
Five real arenas from the announcement. Each one is a playable micro-challenge — beat the clock the way Ultrafast does, and bank the points.
ARCADE
Ride the interconnect fabric. Harvest SOL tokens and cache bursts; dodge thermal spikes. The fabric speeds up — like demand on a wafer-scale engine.
Works with mouse, touch, and keyboard. Every token you harvest is one the GPU never saw coming.
CERTIFICATION
Eight questions. Streaks multiply. Finish certified — or finish Standard.
DEPLOYED WITH
“The increase in speed brought by Cerebras is impressive. It enables different ways of using the models, and makes it practical for developers to work in a more focused and productive way alongside them.”
OpenAI runs Ultrafast on its own fires: incident response that reads logs, analyses traces and synthesises conversations while evidence is still changing — and research loops tightened from overnight batches into multiple same-day iterations.
Ultrafast mode is the next step in the OpenAI–Cerebras partnership to bring ultra-low-latency inference to OpenAI’s platform — Cerebras now supporting OpenAI’s most intelligent model.
Limited preview today for a select group of API customers, expanding as capacity grows. OpenAI is collecting access requests through a sign-up form linked from the announcement.
READ THE ANNOUNCEMENT ↗