OpenAI previews Ultrafast tier for GPT-5.6 Sol

OpenAI has unveiled an invite-only preview of Ultrafast, a new API tier designed to dramatically accelerate its flagship GPT-5.6 Sol model. Powered by a partnership with hardware developer Cerebras, the new tier can generate outputs up to 14 times faster than standard speeds, reaching processing rates as high as 750 tokens per second. The infrastructure behind this speedup stems from a collaboration announced in January, which outlines plans to deploy 750MW of Cerebras's specialized, speed-focused compute.

To demonstrate the real-world capabilities of the accelerated model, OpenAI tested Sol with Ultrafast on Humanity's Last Exam. The system completed the 2500-question test in just 11 hours, a massive reduction compared to the 78 hours required by the Fable model, while achieving comparable results. Internal feedback from OpenAI employees highlights the practical impact of this performance boost. One staffer noted that the extreme speed made them feel like they were "genuinely cheating at my job," while another reported that security investigations that previously took hours were completed in just 10 minutes.

Currently, Ultrafast is available only as an invite-only API preview, and OpenAI has not yet disclosed any pricing details. The company plans to expand access gradually as more Cerebras compute capacity comes online. For AI practitioners and developers, this shift represents a major evolution in workflow design. Operating at 750 tokens per second allows for the creation of highly complex, multi-agent systems and automated workflows that can reason and respond in near-real-time. This eliminates the traditional latency bottleneck that has long forced developers to choose between frontier-level intelligence and rapid execution.