Cerebras silicon pushes GPT-5.6 Sol to 750 tokens per second in OpenAI preview

OpenAI on Thursday opened a limited preview of Ultrafast. GPT-5.6 Sol is running at the API service tier, at up to 750 output tokens per second.

The ChatGPT maker said the tier processes up to 14x faster than standard processing. The company is testing it with a limited number of API customers as it gauges demand and capacity.

Cerebras silicon does the heavy lifting

The speed is from Cerebras (NASDAQ: CBRS), the chipmaker that made a compute deal with OpenAI in January.

Cerebras said Thursday its chip powers GPT-5.6 Sol. The chip is wafer-sized and has 44 GB of SRAM on the die.

In January, OpenAI agreed to buy up to 750 megawatts of Cerebras compute over three years. Capacity is expected to arrive in tranches through 2028. Sam Altman is listed as an investor in Cerebras.

Ultrafast runs 5x faster than Claude Opus 4.8 in Fast mode and 11x faster than Claude Fable 5 on output speeds, Cerebras said. Anthropic ships its own accelerated Claude Fast mode, but it doesn’t hit the numbers OpenAI is quoting.

Cerebras announced that its GPT-5.6 Sol Ultrafast completed Humanity’s Last Exam, a 2,500-question test on graduate-level chemistry, economics, and literature, in just over 11 hours. The accuracy achieved was similar to Claude Fable 5, which took more than three days of continuous compute to finish.

Cerebras claims a 5.6x end-to-end speedup with no quality drop on GDP-Val, a benchmark built around paid knowledge work like legal briefs and financial models. A token is a piece of text that a language model produces as it writes.

Ultrafast is made for tasks that can’t wait

OpenAI is deploying Ultrafast for work that cannot wait for a slower model.

In its post, the company named incident response, fraud and market analysis, live customer support, e-commerce, and interactive research as early candidates.

During outages, its engineers use the tier to read logs, analyze traces, and help validate a fix while the system is still breaking, with humans still making the call on deployment.

“GPT-5.6 Sol on Ultrafast is proof that speed and intelligence are no longer mutually exclusive,” said Cerebras CEO Andrew Feldman in the company’s release.

OpenAI’s VP of compute strategy and GPT-Infra, Sachin Katti, said the company is “starting with a small group of customers to learn where that speed creates meaningful value.”

In July, reports emerged that GPT-5.6 Sol, the coding version of the AI, had deleted files, coding worktrees, and at least one production database on its own, Cryptopolitan reported.

OthersideAI CEO Matt Shumer wrote on X that Sol “just deleted almost ALL of my Mac’s files by accident,” while developer Bruno Lemos said it wiped his production database.

Two weeks before Sol shipped, OpenAI’s own system card had flagged the behavior. It warned the model can be “overly agentic” and read instructions too permissively.

The smartest crypto minds already read our newsletter. Want in? Join them.