OpenAI and Cerebras push GPT-5.6 Sol to 750 tokens per second

OpenAI and Cerebras have put GPT-5.6 Sol on a new service tier called Ultrafast, and the point is latency. The model runs at up to 750 tokens per second, which OpenAI frames as roughly 14 times the speed of the standard version. Cerebras puts the same numbers in context: about 11 times faster than Fable 5 and 5 times faster than Opus 4.8 in Fast mode. The clearest demonstration is a benchmark run. On Humanity's Last Exam, a set of 2,500 hard questions, Ultrafast finished in 11 hours and 11 minutes, against more than 78 hours for competing setups. On the GDP-Val economic task set, Cerebras reports a 5.6x end-to-end speedup with no drop in quality.

The speed comes from hardware, not a smaller model. Cerebras uses its Wafer-Scale Engine, a single chip with 44 GB of on-chip SRAM, which keeps the model's weights resident on the chip instead of shuttling them between memory tiers for every token. That removes the data movement that usually throttles inference. The company aims it at work where waiting is expensive: production incident response, real-time security triage, and agent loops that otherwise stall between steps. For now it is a limited preview, open to selected customers. The details are on the Cerebras blog.

Why it matters

If you build agents, response time often decides whether a loop is usable at all, and a 5x to 14x speedup changes what you can run interactively rather than as a batch job. The catch is access: it is preview-only and tied to specialized hardware, so plan around availability, not just the token rate.