News · 2026-08-13

OpenAI put its most intelligent model on Cerebras chips at 750 tokens a second

OpenAI is previewing Ultrafast, a new service tier that runs GPT-5.6 Sol up to 14 times faster than standard processing and generates up to 750 output tokens per second. It is powered by Cerebras hardware and launches first in the OpenAI API. The announcement, dated August 13, 2026, is explicit that this is a serving change and not a new model: the same frontier model, on different silicon, answering roughly an order of magnitude faster.

Key facts

• Up to 14 times faster than standard processing, and up to 750 output tokens per second.
• Runs GPT-5.6 Sol, OpenAI's most intelligent model, with no advertised capability reduction.
• Limited preview to a select group of API customers as of August 13, 2026, expanding as capacity grows. No separate Ultrafast price is published.
• Primary source: OpenAI, Previewing Ultrafast mode, August 13, 2026.

The trade OpenAI is claiming to have broken is an old one. Until now, as the company puts it, "getting real-time speed typically meant choosing a smaller or more specialized model." Every product team that wanted an answer inside a live conversation had to drop down a tier and accept a dumber model. Ultrafast points, in OpenAI's words, "in a new direction: more useful work per second."

The reason this needs special hardware is worth understanding, because it is the single most counterintuitive fact about running language models. Generating text is not limited by how fast a chip can multiply. It is limited by how fast the model's weights can be dragged out of memory and into the processor, one token at a time -- a constraint explained in full in our lesson on why AI inference runs out of memory bandwidth before it runs out of math. A conventional accelerator spends most of its time waiting on memory. Cerebras builds a single wafer-sized chip that keeps model weights in on-chip memory rather than fetching them across an external bus, which attacks exactly that bottleneck. The analogy is a chef who keeps every ingredient on the counter instead of walking to the pantry for each one: the cooking was never the slow part.

Early customers describe the change as qualitative rather than incremental. "The increase in speed brought by Cerebras is impressive," said John Crepezzi of AI Assistants at Jane Street. "It enables different ways of using the models, and makes it practical for developers to work in a more focused and productive way alongside them." OpenAI lists the workloads it thinks are unlocked: incident response while an outage is still unfolding, fraud and market analysis while conditions are still changing, voice support that resolves multi-step problems without a pause in the conversation, and commerce questions answered before a shopper abandons a cart.

The most telling example is OpenAI's own. Its research teams describe a standard rhythm of launching a batch of experiments overnight and reading results in the morning. With Ultrafast, the company says, "we see this loop tightening to support multiple iterations during the workday instead." That is the real product here. Speed past a certain threshold does not make an existing workflow faster; it deletes the workflow and replaces it with a different one.

Two clarifications, because the naming invites confusion. Ultrafast is not "ultra," the reasoning mode OpenAI introduced with GPT-5.6 that coordinates several agents across parallel workstreams. One is about how hard the model thinks; the other is about how fast the same thinking is delivered. And Ultrafast is not a distilled or quantized variant. The announcement makes no claim of reduced capability, which distinguishes it from the usual route to low latency, where speed is bought with distillation, quantization, or speculative decoding and paid for in quality.

The honest caveat is what OpenAI has not published. There is no Ultrafast price in the announcement, no throughput guarantee, no context-length or feature-parity statement, and no independent latency measurement. Access is restricted to a select group of customers, with OpenAI saying only that it will expand "as capacity grows" -- which is a capacity statement, not a schedule. The base model's published API price remains what it was; whether a 14x speedup carries a premium is, for now, unknown to anyone outside the preview.

It also lands in a market where inference cost and speed have become the main competitive surface rather than raw capability. In the same week, DeepSeek began charging rush-hour rates to shape demand across the day rather than expand capacity, and a new agent benchmark found that the two top models scored within a point of each other while differing sharply in cost and token usage. Frontier intelligence is increasingly a commodity with a delivery schedule attached, and the schedule is where the differentiation now sits.

Primary source, verified: read the paper →

Ground Truth. (2026, August 13). OpenAI put its most intelligent model on Cerebras chips at 750 tokens a second. Ground Truth. https://groundtruth.day/news/openai-put-its-most-intelligent-model-on-cerebras-chips-at-750-tokens-a-second.html

@misc{groundtruth:openai-put-its-most-intelligent-model-on-cerebras-chips-at-750-tokens-a-second,
  title  = {OpenAI put its most intelligent model on Cerebras chips at 750 tokens a second},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/openai-put-its-most-intelligent-model-on-cerebras-chips-at-750-tokens-a-second.html}
}

Topics: openai · inference · hardware · cerebras · api · latency