OpenAI began a limited preview of Ultrafast, a new processing tier that runs its GPT-5.6 Sol model on inference chips built by Cerebras Systems rather than the graphics processors that power the rest of the API. The tier generates up to 750 output tokens per second, as much as 14 times the pace of OpenAI’s standard processing, according to the company. The hardware swap is the whole story: OpenAI is not shrinking the model to get the speed, it is renting different silicon.
That distinction matters because the usual fix for latency in AI products is a smaller, faster model, accepted along with the quality it gives up. Voice agents, live support tools, and real-time trading systems have generally had to choose between a frontier model that answers well and a lighter one that answers fast. Ultrafast is OpenAI’s bet that a chip built specifically for fast token generation can substitute for that compromise, keeping the full Sol model while buying the speed elsewhere.
Access is not open. OpenAI is running the preview with an initial set of customers spanning coding, commerce, financial research, and customer support, and it frames the limit as one of capacity rather than product readiness: the company says it will widen the group only as more Cerebras capacity comes online. Ultrafast currently exists only inside the OpenAI API. There is no consumer surface, no ChatGPT toggle, and no stated date for general availability.
Among the named early customers is Jane Street, the trading firm. John Crepezzi, who works on AI assistants at Jane Street, said in OpenAI’s announcement: “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 also named Podium, Basis, and Rogo as early access customers, without disclosing usage volumes or contract terms for any of them.
OpenAI is pitching the tier at workloads that previously had to be interrupted, queued, or run overnight:
- Incident response, reading logs and recent code changes to help identify a cause while an outage is still active
- Financial research and fraud checks against market data that keeps changing in real time
- Customer support conversations that need several lookups without breaking the flow of the call
- Commerce questions about inventory, pricing, or checkout issues while a shopper is still deciding
Inside OpenAI, engineers have reportedly used the mode for their own incident response, and the company says a research workflow that used to mean launching experiments overnight and checking results the next morning is starting to compress into multiple iterations across a single workday. Neither claim comes with measured before-and-after figures in the announcement.
OpenAI describes Ultrafast as the next step in its partnership with Cerebras on low-latency inference, not a one-off feature. That framing points to a structural choice: OpenAI is diversifying which chips serve its models rather than treating Nvidia GPUs as the only production path for frontier-scale inference. Cerebras builds wafer-scale processors designed around fast, low-latency generation, a different tradeoff than the general-purpose GPUs the rest of OpenAI’s traffic runs on.
If Ultrafast holds its speed and quality once it exits preview, the tradeoff that has shaped a generation of AI products, smaller and faster versus larger and slower, stops being fixed. Teams building voice agents, trading tools, or live support systems could keep a frontier model in the loop and pay for latency with hardware instead of capability. Until access widens beyond a hand-picked customer list, though, that is a design assumption to plan around, not something to build into a production system today.
OpenAI detailed the Ultrafast preview in a company announcement published August 13, 2026.
