OpenAI's Ultrafast Tier Puts Its Flagship Model on Cerebras, Not GPUs

OpenAI has put its flagship model on someone else's silicon. On August 13 the company opened a limited preview of "Ultrafast mode" for GPT-5.6 Sol, quoting up to 750 output tokens per second and up to 14 times the speed of the standard version of the same model, with the tier powered by Cerebras hardware rather than conventional GPUs ^[1][2][3]. The consequence is not a speed record; it is that latency has become a line item you buy rather than a property of the model you chose.

Start with the arithmetic, because the marketing multiple hides it. If Ultrafast delivers 750 tokens per second at 14x the standard tier, the standard tier is serving roughly 54 tokens per second ^[9]. The same source says Ultrafast runs 11x faster than Fable 5, OpenAI's previous-generation model, and 5x faster than a model it names only as Opus 4.8 in Fast mode ^[4], which implies about 68 and 150 tokens per second respectively ^[10]. Read those together and the default serving speed of the new flagship looks slower than the previous generation's ^[11]. Ultrafast is not new headroom so much as a way to buy back the latency the bigger model gave away.

The pitch is aimed at enterprise workloads where the pause is the product problem: real-time voice, financial research, security response ^[5]. The publication's framing goes further, suggesting a general-purpose model at these speeds could replace or augment the bespoke NLP systems hedge funds and trading desks use to parse earnings calls and filings ^[5]. That is a claim about substitution economics, and the source gives no price for the Ultrafast tier, so nobody outside the preview can test it ^[13]. A page of text per second is useful ^[2]; whether it is cheaper than the pipeline it displaces is unstated.

The rollout terms are the tell. Access is limited to select customers through the OpenAI API, with broader availability planned once capacity scales up ^[8]. Software features do not wait on capacity. Wafer-scale silicon does, and Cerebras builds processors that dwarf conventional GPUs ^[3]. GPT-5.6 Sol launched in July 2026 and the Ultrafast preview arrived barely a month later ^[6][7], which suggests the constraint on shipping it was supply rather than engineering.

For anyone buying inference, three things change. First, the fast tier and the cheap tier may run on different hardware, so region availability, rate limits and reliability characteristics stop being uniform across a single model name. Second, latency benchmarks quoted against a vendor's own default tier tell you little about the market; the useful comparison is against whatever you run today. Third, a frontier model served at scale on non-GPU hardware means inference capacity is being sourced outside the GPU supply chain, though the reporting, which cryptobriefing.com credits to gizmodo.com, names no dollar figure, no contract term and no other chip supplier ^[12][14][13].