GPT-5.6 Sol gets an API speed tier, but only a limited preview

OpenAI has introduced Ultrafast, a new OpenAI API service tier for GPT-5.6 Sol. The company says the tier can run Sol up to 14 times faster than Standard processing and generate up to 750 output tokens per second through Cerebras. The change is a serving option, not a newly named model: OpenAI says access is currently limited to a select customer group while capacity grows. That makes the immediate story about latency and availability rather than a change to the model's published capabilities.

The timing matters because OpenAI published a separate builder guide on the same day. That guide presents GPT-5.6 as a family meant to reduce agent cost and overhead through lower reasoning effort, persisted reasoning, compaction, native multi-agent orchestration, programmatic tool calling, and a longer prompt-cache lifetime. Ultrafast therefore sits beside changes to how an application can organize work. The company is positioning faster generation as useful when an interactive system must complete several model and tool steps before the user loses context.

The comparison with GPT-5.5 needs restraint. OpenAI reports that Sol at low reasoning outperformed GPT-5.5 at high reasoning on Agents' Last Exam when the harness was held constant, but that is a provider-reported result rather than an independent benchmark in these sources. Axios described the original GPT-5.6 rollout in June as restricted access during a government review process. The current Ultrafast announcement also describes a select-customer preview, but identifies capacity growth, not government review, as the stated condition for expansion. OpenAI has not published Ultrafast pricing or a service-level commitment in the announcement.

For technical teams, the relevant measurement is end-to-end task time, not the headline token rate alone. The likely fit is a workflow with repeated model turns and tool calls, such as incident triage, research, support, or a live coding loop, where wait time compounds. Teams should compare completion time, cost, reliability, and quality against Standard processing on their own workload before redesigning an agent around the tier. The availability limit and absence of public pricing mean the announcement is evidence of a new deployment option, not evidence that the option is ready for broad production planning.