OpenAI introduced Ultrafast — a mode in which GPT-5.6 Sol operates 14 times faster

OpenAI has launched a limited preview of Ultrafast — a new tier of service in its API, where the flagship GPT-5.6 Sol model processes requests up to 14 times faster than the standard mode. The acceleration is powered by Cerebras chips, and the announcement was made on August 13.

Ultrafast is not a new model, but a separate pricing tier running the existing GPT-5.6 Sol. The company reports that the mode delivers up to 750 output tokens per second — tokens being fragments of text that a language model generates during operation. The model's intelligence, response quality, and context window remain unchanged; only the speed at which tokens are delivered changes. In its blog, OpenAI notes that until now, achieving near-real-time speeds typically meant switching to a smaller or more specialized model, whereas Ultrafast sets a different direction — more useful work per second.

The company identifies five scenarios where speed is critical: incident response and system reliability, financial research and transaction security, customer support and voice assistants, e-commerce, and live research and experiments. Within OpenAI itself, Ultrafast is used during incident analysis — the model reads logs, analyzes traces, and helps prepare fixes while the outage is still ongoing. The preview is already being tested by companies in development, commerce, financial research, and support sectors, including Jane Street, Podium, Basis, and Rogo.

The acceleration marks the next step in OpenAI’s partnership with Cerebras, aimed at ultra-low-latency inference. GPT-5.6 Sol in Ultrafast mode runs 11 times faster than Fable 5 and five times faster than Opus 4.8 in Fast mode. Anthropic has its own Fast mode, but it does not reach the performance levels claimed by OpenAI. Ultrafast is available to a narrow group of clients and works only via the API; it is not available in ChatGPT. Access will expand as capacity grows. The pricing for the tier, the date of general availability, and a separate model identifier have not yet been announced.

Source: OpenAI