OpenAI published Ultrafast on August 13: a new API service tier that runs its flagship GPT-5.6 Sol model up to 14 times faster than standard processing. The company says the stack, built with Cerebras, can emit as many as 750 output tokens per second. That is meant to keep Sol’s quality while removing the usual trade — if you wanted real-time speed, you used a smaller model.
Ultrafast is not in ChatGPT for everyone. It is a limited API preview for a first group of customers, with a signup list for later access. Early testers named in the post include Jane Street, the voice-AI team at Podium, Basis, and Rogo. OpenAI’s own examples are incident response (read logs and traces while an outage is still moving), live financial research, multi-step customer support and voice, checkout help before a shopper leaves, and research loops that used to run overnight.
Inside OpenAI, a developer group is already using the tier to shrink the gap between seeing an alert and testing a fix. The company is careful to say engineers still own the judgment and the deploy. Cerebras is described as the next step in an existing inference partnership, now carrying OpenAI’s most capable model rather than a cut-down one.
No public price list was in the launch post. The claim to weigh is simple: frontier-class answers at interactive speed, for the customers who get in.
Source: https://openai.com/index/previewing-ultrafast/