OpenAI Ultrafast: Limited Preview, 14× Faster

In preview and reserved for a small group of clients, OpenAI's Ultrafast mode relies on a partnership with chip manufacturer Cerebras. OpenAI promises, for GPT-5.6 Sol, up to 14 times the usual speed and up to 750 tokens of output per second, with expanded access announced as capacity increases. The company presents this mode as a pathway to more useful work per second and mentions enterprise deployments.

Restricted Access and Support from Cerebras for the Preview

Ultrafast is currently offered in a preview phase and is only accessible to a small group of clients. OpenAI states that it will extend access as capacity increases. This mode relies on a partnership with chip manufacturer Cerebras.

Announced Performance: 14× and 750 tokens/s for GPT-5.6 Sol

OpenAI has launched Ultrafast, a mode designed to accelerate the work of GPT-5.6 Sol, presented by the company as its most powerful model. According to OpenAI, this mode can achieve 14 times the standard processing speed and produce up to 750 tokens of output per second. A token corresponds to a unit of text generated by a language model during an interaction. The company claims that achieving real-time speeds previously required choosing a smaller or specialized model, and sees Ultrafast as another pathway to "more useful work per second." On the competition side, Anthropic also offers accelerated versions of its models: Claude has a fast mode, which does not display the same speed as that announced by OpenAI.

Proposed Deployments in Four Business Functions

OpenAI suggests that this high-performance version of GPT-5.6 Sol can be integrated into workflows such as incident response, customer service and support, financial market analysis, or e-commerce. These usage scenarios are put forward by the company as part of the preview.