The tier announced

OpenAI has opened a preview of Ultrafast, a mode that runs the GPT-5.6 Sol model up to 14 times faster than standard processing. Offered first through the OpenAI API, the new service tier can produce up to 750 output tokens per second.

The stated goal is to reduce the need to reach for smaller or task-specialised models purely to get a faster response. In other words, the choice is meant to stop being a trade-off between speed and capability: OpenAI says latency is substantially reduced while GPT-5.6 Sol's capabilities are preserved.

The infrastructure side

The infrastructure behind the new mode is provided by Cerebras, which develops AI chips. In that respect Ultrafast is not only a service tier but a new stage in the low-latency inference collaboration between OpenAI and Cerebras.

That distinction matters technically. 750 tokens per second is a generation rate that is hard to reach on general-purpose GPU clusters; figures at that level are typically associated with hardware specialised for inference. Ultrafast is therefore the result of a choice made not on the model side but on the hardware the model runs on.

Where it is used

OpenAI positions Ultrafast for areas where speed is critical:

  • Incident response: application logs and code changes can be examined in real time during system failures.
  • E-commerce: stock checks, product recommendations and problems in the payment flow can be handled before the user abandons the purchase.
  • Financial research, security, customer support and voice AI applications: other areas where response time feeds directly into the experience.

The company's own developers use the mode to examine logs and traces during system incidents. On the research side, experiment runs that previously continued overnight can now be repeated several times within the same working day.

Access and unknowns

Ultrafast is currently offered to a limited group of users, which includes companies working in coding, financial research, e-commerce and customer support. OpenAI plans to widen access as capacity increases. The company has not yet announced pricing for the mode or when it will become generally available.

Why it matters

The significance of the announcement is that it touches a habit long established in the industry. When response speed is critical, teams generally switch to a smaller and weaker model, so latency falls but accuracy falls with it. Ultrafast's promise is to remove that trade-off.

Testing that promise, though, requires two pieces of information. The first is price: how much does the speed gain cost per token? The second is capacity: whether a tier resting on specialised hardware can hold the same speed once widely available can only be seen as access expands. That the company names capacity as the condition for widening access suggests that limit is understood.

Another question is which workloads genuinely benefit from that speed. In an application producing a single long response, tokens per second translates directly into waiting time; but in multi-step agent flows the total is often set by the latency of tool calls and external services. How a claimed 14-fold speed-up shows up in the end-to-end experience will therefore vary markedly with the shape of the application.