OpenAI's new Ultrafast mode makes GPT-5.6 Sol run at 14 times normal speed

Key points

• OpenAI launched a new mode called Ultrafast for its GPT-5.6 Sol model on 13 August 2026.
• Ultrafast generates up to 750 tokens (pieces of text) per second, which OpenAI says is 14 times faster than standard processing.
• The feature runs on chips from Cerebras, a specialist AI chip company, under an existing partnership with OpenAI.
• Ultrafast is currently in preview and available to a limited group of business customers only.
• Anthropic's Claude chatbot has a similar fast mode, but OpenAI claims its speed figures are higher.

If you have ever typed a question into ChatGPT and drummed your fingers while it thought, OpenAI has noticed.

The company announced Ultrafast, a new speed mode for GPT-5.6 Sol, its latest and most capable model, on Thursday. The promise: responses that arrive 14 times faster than usual, generating up to 750 tokens per second. A token, in plain terms, is a small chunk of text, roughly three quarters of a word, that a large language model (the technology behind chatbots like ChatGPT) produces as it builds a reply.

To put that in human terms: a typical assistant might read a message and draft a reply in the time it takes you to blink. Ultrafast is aiming to make that feel almost instant, even for long outputs.

Who is this actually for?

Right now, not the average ChatGPT subscriber. OpenAI is rolling out Ultrafast as a preview for a small group of business customers, with plans to expand as capacity grows.

The company is pitching it squarely at corporate use cases where speed genuinely matters: incident response (catching a server outage before it spreads), customer service queues, financial market analysis, and e-commerce.

"Until now, getting real-time speed typically meant choosing a smaller or more specialized model," OpenAI wrote in its announcement. "Ultrafast points to progress in a new direction: more useful work per second."

That quote matters. Faster AI has historically meant simpler AI. Ultrafast, if the claim holds, would let businesses get speed without trading away quality.

What makes it so fast?

The speed comes from a chip partnership. Ultrafast runs on hardware from Cerebras, a company that builds specialised processors designed for exactly this kind of heavy AI workload, rather than the general-purpose chips found in most computers.

OpenAI's main rival Anthropic, the company behind the Claude chatbot, offers its own fast mode. OpenAI's 750-tokens-per-second figure is a higher claim than Claude's current fast mode delivers, though direct comparisons depend on what task you throw at either model.

What does this mean for everyday users?

Nothing to do right now. Ultrafast is a business-tier, invite-only preview. Regular ChatGPT users will not see it in their settings today.

If you work somewhere that already uses OpenAI's API (the programming interface companies use to plug ChatGPT into their own products), it is worth watching for the wider rollout. For everyone else, the practical takeaway is that response times across AI tools are improving fast, and the competition between OpenAI and Anthropic is likely to push that even further.

Common questions

Will Ultrafast cost more than standard ChatGPT?

OpenAI has not published pricing for Ultrafast yet. Speed-optimised tiers in AI services typically carry a higher per-use cost for businesses, so expect a premium once it leaves preview.

Does Ultrafast change what GPT-5.6 Sol can do, or just how fast it does it?

Speed only, based on what OpenAI has shared. The underlying model is the same; Ultrafast is about delivering its answers faster, not giving it new abilities.

Is my data handled differently in Ultrafast mode?

OpenAI has not announced separate privacy terms for Ultrafast. The same data-handling rules that apply to GPT-5.6 Sol in standard mode should apply here, but businesses should check their enterprise agreements before rolling it out.