OpenAI's Ultrafast mode pushes GPT-5.6 Sol to 14x speed in a preview

OpenAI has rolled out a new mode called Ultrafast, aimed at making its current most powerful model, GPT-5.6 Sol, work at a much faster pace. The company says the mode runs at 14x the speed of standard processing and can produce up to 750 output tokens per second.

That token rate is the hard number in the announcement. Standard generation is measured in tokens per second, and 750 is high for a frontier model in a normal run. OpenAI frames it as a way to get more useful work out of each second, rather than trading quality for speed.

"Until now, getting real-time speed typically meant choosing a smaller or more specialized model," the company said in the blog post that announced the mode on Thursday. "Ultrafast points to progress in a new direction: more useful work per second."

What it is for

OpenAI points to a set of business use cases where latency actually matters: incident response, customer service and support, financial market analysis, and e-commerce. The common thread is that the answer has to be there fast, not just correct.

The move follows a pattern competitors have already started. Anthropic, for example, ships accelerated versions of its models. Claude has a fast mode, though it does not reach the throughput OpenAI is claiming here.

Who can use it

Ultrafast is currently in preview. It runs on a partnership between OpenAI and chipmaker Cerebras, and for now access is limited to a small group of customers. The company says it will widen availability "as capacity grows."

For a team deciding whether this changes anything, the practical read is narrow at this stage. The speed claim is a vendor number, and the preview is closed to most users, so there is nothing to benchmark against in production yet. If you run latency-sensitive workloads such as incident response or live customer support, it is worth watching, since that is where 750 tokens per second would actually show up in the bill and in response time. If your work is batch or offline, standard GPT-5.6 Sol is the more relevant model to cost against, and the speed premium buys you little.