On 13 August OpenAI previewed Ultrafast mode, a new service tier for GPT-5.6 Sol that runs the model up to 14 times faster than the standard endpoint, peaking at around 750 output tokens per second. It is not OpenAI’s own silicon doing the work: the speed comes from Cerebras, whose wafer-scale chips trade GPU-style batching for raw, single-request throughput.
Standard GPT-5.6 Sol prices at $5 per million input tokens and $30 per million output tokens as of this week — worth stating because API prices move fast, and Ultrafast doesn’t have a public rate yet. It’s a waitlist preview: limited customers, no quota, no region list, no general-availability date.
Why wafer-scale changes the shape of the problem
A GPU serving an LLM is optimised to answer many requests at once, cheaply, in parallel. That’s the right trade-off for a chat product with millions of users typing at their own pace. It’s the wrong trade-off the moment one request needs to finish before something else can happen — a tool call, a render, a decision a person is sitting and waiting on.
Cerebras’ Wafer-Scale Engine keeps an entire model’s weights resident on one enormous chip instead of splitting them across a cluster and shuttling activations between cards. Less communication overhead per token means a single stream can move much faster, at the cost of being a worse fit for serving thousands of streams at once. OpenAI is explicitly pointing Ultrafast at the cases where that trade makes sense: financial research, incident response, customer support, voice, commerce, live experimentation — anywhere the model is one link in a chain that a person or another system is blocked on.
Where this actually lands in a pipeline
I don’t reach for a chatbot mid-shot. What I do reach for, constantly, is a script that renames three hundred objects, a query against an asset naming convention, a quick sanity check on why a shader graph is throwing a warning. Right now, any of that routed through an LLM has a beat of dead air in it — small, but it’s a beat that breaks flow, so you end up batching those questions instead of asking them as they come up.
750 tokens a second is fast enough to close that gap. Not fast enough to replace judgement — nothing here changes what actually doesn’t delegate — but fast enough that a tool bolted into a DCC app, a build pipeline, or a bug-triage pass during review could answer back inside the same breath you asked the question in, instead of you tabbing away and losing the thread.
Voice is the other one worth watching. A model that replies at 750 tokens a second is fast enough for a spoken back-and-forth that doesn’t feel like talking to an answering machine — useful the moment someone wires a voice layer onto a review tool or a scene-navigation assistant, hands busy with a tablet or a controller.
Worth knowing about, not worth building on yet
This is a preview behind a waitlist, with no published price and no committed capacity. None of that makes it a fad — Cerebras has been shipping fast inference for a while, and OpenAI putting a flagship model on it is a real signal about where they think the demand is. It just means the honest read today is “interesting, watch the general-availability terms,” not “go build your pipeline around it.” I’ll believe the throughput numbers once they’re generally available and I can point a real workflow at them myself.