OpenAI is running GPT-5.6 Sol on Cerebras wafer-scale chips at up to 750 tokens per second

OpenAI made the GPT-5.6 family generally available on July 9, moving its flagship Sol model, along with the cheaper Terra and Luna tiers, out of the roughly twenty-company preview it had run since late June and into ChatGPT, Codex, and the API. The rollout is global and phased over about a day.

The model card is not the interesting part. The interesting part is one sentence OpenAI buried at the bottom of its June preview note and has now put into production: Sol is running on Cerebras hardware at up to 750 tokens per second.

What 750 is measured against

A raw throughput number means nothing without the baseline. Here is the baseline. A typical Nvidia H100 cluster serving a frontier model in production generates on the order of 70 tokens per second. Cerebras is quoting up to 750. That is where the "roughly ten times faster" framing comes from, and it is a production deployment against a production deployment, not a single-chip benchmark run under lab conditions.

Two caveats belong next to that number, and OpenAI supplies both. The first is the phrase "up to," which is doing work. The second is a footnote OpenAI attaches to its own latency figures: they are estimated by simulating production behavior offline, and "real-world results may vary substantially." So treat 750 as a ceiling the vendor is comfortable advertising, not a floor you will hit on every request.

Why the speed is worth caring about

For a chatbot answering one question, 750 tokens per second versus 70 is the difference between fast and slightly faster. Nobody reads that quickly. The reason this matters is agents.

An agentic workflow chains many generations back to back: the model writes a tool call, waits for the result, reads it, writes the next one, and repeats, sometimes for dozens of steps. The total time is dominated by how fast the model emits tokens, because the steps are sequential. At 70 tokens per second a multi-step task that runs for 30 seconds is a workflow people abandon. At 750 the same task finishes in around three seconds, which is the range where people leave it switched on. Speed stops being a spec and starts being the thing that decides whether the product gets used.

Why a wafer does this and a GPU struggles

Token generation is memory-bound, not compute-bound. To produce each new token the chip has to read the model's weights, and on a GPU those weights sit in high-bandwidth memory next to the die. The rate you can stream them out of HBM caps how fast you can generate. Adding more compute does not help if the memory pipe is the bottleneck.

Cerebras builds one chip the size of a wafer and keeps the weights in on-chip SRAM spread across it, which has far more bandwidth than HBM. That is the mechanism behind the throughput. It is real engineering, and it is why wafer-scale parts post token-generation numbers that GPU clusters do not.

The tradeoff is that a wafer is enormous and expensive, and a frontier model does not fit on one. It has to be split across many. An estimate circulating from developer Bleys Goodson, based on the quoted speed, puts Sol across roughly 70 to 100 wafers with about one model layer per wafer. That figure is inference from the outside, not a disclosed spec, so hold it loosely. But the direction is right, and it explains OpenAI's own hedge: access to the Cerebras endpoint is "initially limited to select customers as we expand capacity." Silicon that large is supply-constrained, and dedicating scores of wafers to a single model replica is not something you switch on for everyone in a week.

The commercial read

Cerebras is marketing its S-1 ahead of a planned 2026 IPO. Landing OpenAI's flagship as a reference deployment weeks before that roadshow is useful timing, and worth naming plainly rather than pretending the engineering exists in a vacuum. The competitive context: Groq serves Meta's Llama models at around 500 tokens per second, and SambaNova runs older Llama models nearer 250. None of them had booked an OpenAI frontier model. Cerebras now has, and it is the first time an OpenAI frontier model has run at production scale on something other than Nvidia.

For pricing, Sol lands at $5 per million input tokens and $30 per million output on the standard API. Terra is $2.50 and $15, Luna $1 and $6. Whether the Cerebras speed tier carries a premium, or Cerebras eats the cost to win the logo, is the number to watch next.