Same model, same weights, different silicon: the new API tier parks GPT-5.6 Sol in on-chip SRAM and skips the memory bus entirely.
OpenAI has opened a limited preview of OpenAI Ultrafast, an API service tier that serves GPT-5.6 Sol at up to 750 output tokens per second, roughly 14 times what the Standard tier manages. The model itself has not changed. The hardware has: Ultrafast requests run on Cerebras wafer-scale systems instead of the GPU fleet behind everything else the company ships.
Why a wafer beats a GPU at this one job #
Generating tokens one at a time is a memory problem dressed up as a compute problem. Every token requires pulling the model weights through the arithmetic units, and on a conventional accelerator those weights sit in HBM stacked beside the die. The math finishes early and waits on the bus. Piling on FLOPS does not help; bandwidth sets the ceiling.
Cerebras deletes the trip. Its Wafer-Scale Engine carries 44GB of SRAM on the die itself, enough to hold the weights where the multipliers live, so nothing has to be streamed in from off-package memory while the model is decoding. On-die SRAM is an order of magnitude quicker than HBM and it never leaves the chip. That is an absurd way to build silicon if you care about capacity per dollar, and close to ideal if you care about how fast a single conversation moves.
The numbers #
OpenAI is insistent that this is not a distilled or quantised variant: Ultrafast and Standard run the same GPT-5.6 Sol. Working through the 2,500 questions of Humanity’s Last Exam, Ultrafast finished in 11 hours and 11 minutes; Cerebras says Claude Fable 5 needed 78 hours and 27 minutes on the same set, close to seven times longer. On GDP-Val, OpenAI reports a 5.6x end-to-end speedup over Standard with no measurable quality loss, and claims roughly a 5x lead over Claude Opus 4.8 running in Fast mode.
Vendor benchmarks deserve the usual pinch of salt. Each of those figures comes from a serving configuration the vendor controls, and 750 tokens per second is a peak rather than a promise. The gap is wide enough, though, that nobody is arguing about measurement error.
Who actually needs 750 tokens a second #
Nobody reads that fast, which is rather the point. The tier is aimed at work where the model is not talking to a human in real time: voice pipelines, support automation, developer agents that chain dozens of calls before producing anything visible, financial research, security response. Reasoning models burn tokens thinking before they answer, and when a task involves twenty sequential calls, token rate turns directly into wall-clock time. A five-minute agent run becomes a thirty-second one.
Ultrafast is a waitlisted preview for now, open to a small group of API customers, and pricing has not been detailed publicly. The more interesting detail may be the strategic one. OpenAI is serving a frontier model in production on silicon that is not Nvidia’s, and the reason is not price or supply. It is that the architecture happens to suit the job.