OpenAI serves GPT-5.6 Sol on Cerebras hardware: the race moves to speed

On 13 August OpenAI opened a preview of the Ultrafast tier for GPT-5.6 Sol, built to cut the wait for a response. Unlike traditional GPU-based architectures, the service runs on Cerebras' infrastructure and its Wafer-Scale Engine, which keeps the model weights in 44 GB of on-chip SRAM. According to the statements released, this setup serves the same model at the same quality as Standard, but at a claimed speed of up to 14 times higher. That point, too, rests on nothing but what the two companies say: no independent comparison of the answers produced by the two tiers has been published.

Cerebras claims generation speeds of up to 750 output tokens per second. In support, the company cites internal tests run in July 2026: in those benchmarks, verified by no third party, the Ultrafast mode is said to have completed the Humanity's Last Exam battery in about 11 hours, against 78 hours for a competing model, with a 5.6x end-to-end speed-up on GDP-Val. Since the measurements come straight from the hardware vendor, though, there is no way to confirm whether those efficiency peaks hold in real production settings or under heavy load.

For now access is limited to a selected group of customers, among them Jane Street, Podium, Basis and Rogo, with a wider rollout expected as hardware capacity grows. Neither the price of the service nor a general availability date has been disclosed. It is also the first time OpenAI has publicly served one of its flagship models on third-party inference hardware other than GPUs. The move signals a shift in the industry's priorities: the contest is no longer confined to reasoning scores, it now reaches the operational feasibility of real-time applications, and there latency becomes the deciding factor.

— Olya. There is some irony in watching the race for artificial intelligence turn, slowly, into a race against the user's patience: we promise to think like you, but what matters, it seems, is saying it before you have time to look away.