OpenAI has opened a preview of a sped-up version of its most powerful model, GPT-5.6 Sol. The new inference mode, called Ultrafast, reaches up to 750 output tokens per second — a speed-up of as much as 14x over the standard mode.

The hardware: Cerebras

The speed does not come from a software optimisation but from different silicon. Ultrafast runs on chips from OpenAI's $10 billion partnership with Cerebras.

Cerebras takes a different approach from GPU makers: the company builds an entire silicon wafer as a single chip, an architecture known as wafer-scale. The clearest advantage of that design is that model weights can be held on the chip itself, largely eliminating traffic between memory and processor. In token generation the bottleneck is usually not raw compute but exactly that traffic.

A three-tier structure

Ultrafast does not arrive alone. Alongside Standard and Fast, OpenAI now offers three tiers:

  • Standard — default speed at the default price.
  • Fast — the middle tier.
  • Ultrafast — up to 750 tokens per second, at the highest price.

The real news here is commercial rather than technical. Same model, same output quality, different speed — and a different price. Inference speed becomes a product in its own right.

Why speed matters this much

In a chat interface the difference between 50 and 750 tokens per second means little beyond reading speed, and nobody reads that fast anyway. Where the difference lands is agentic workflows.

Completing a task, an agent makes dozens or even hundreds of model calls: it plans, calls a tool, reads the result, replans. Latency accumulates at every step in that chain. A few seconds saved on a single call becomes minutes across a thirty-step chain.

On the enterprise side that is directly a question of usability. An agent task that finishes in half a minute can sit inside a workflow; the same task taking ten minutes can only run overnight as a batch job.

The competitive angle

Pricing speed separately is an attempt to strengthen OpenAI's position in the enterprise market. As the gap in model quality against rivals narrows, the points of differentiation shift: from quality to latency, and from latency to cost.

The size of the Cerebras partnership supports that reading. A $10 billion commitment suggests speed is being treated as a durable product line rather than a temporary promotion.

The other side of the pricing

The three-tier structure looks like freedom of choice from the user's side, but there is another reading: an arrangement in which the default tier is no longer the best tier. Access to a model used to mean a single price; now how fast that same model runs depends on the budget.

This is a long-familiar pattern in cloud services. Storage and networking have long offered the same data at different access speeds and different prices. AI inference settling into that pattern can be read as a sign of a maturing market — but it also means speed becomes a privilege for smaller developers.

That it is still in preview indicates prices and limits have not settled. Ultrafast's permanent price will be the variable that decides who the mode actually serves.