OpenAI puts latency on the price list: 750 tokens/sec, gated by workload fit

On August 13 OpenAI previewed Ultrafast, a limited-access API service tier that runs GPT-5.6 Sol at up to 750 output tokens per second, described as up to 14 times the speed of Standard processing, on Cerebras hardware ^[1][2][3]. OpenAI presents it as a serving option rather than a new foundation model ^[4], which is the whole point: response speed has moved from a property you inherit when you pick a model to a tier you have to be approved for.

The approval part is not incidental. OpenAI's signup page says capacity is limited and that customer inclusion will be evaluated on workload fit and availability ^[5]. Early testing includes Jane Street, Podium, Basis and Rogo ^[11], and OpenAI names incident response, reliability work, financial research, security, customer support, voice, commerce and live research as candidate workloads ^[6]. Not disclosed in the retrieved material: pricing, regional availability, rate limits, context-window behaviour, and formal service-level commitments ^[9]. There is no general-availability timeline either ^[7]. So the tier currently behaves like allocated capacity with a discretionary gate, and that is a different procurement and architecture problem from choosing between two model names.

Run the arithmetic before designing around the headline. If 750 output tokens per second is 14x Standard, the implied Standard rate is roughly 54 tokens per second ^[20]. Cerebras, which describes its role as powering the service ^[8], reports that on Humanity's Last Exam GPT-5.6 Sol Ultrafast answered 2,500 questions in 11 hours 11 minutes against 78 hours 27 minutes for Anthropic's Claude Fable 5 ^[12]. That is about 7x on wall clock ^[21], half the headline multiple, and it is a cross-vendor comparison run on different dates with different harnesses and reasoning settings ^[13], not a measurement of Ultrafast against Standard. Cerebras also reports a 5.6x end-to-end speedup on GDP-Val without quality degradation and cites Artificial Analysis output-speed data for a 5x edge over Claude Opus 4.8 Fast mode and 11x over Claude Fable 5 ^[14]. Cerebras says the tier delivers the same intelligence as Standard, while OpenAI's public announcement specifies throughput and does not publish independent quality measurements for the new serving configuration ^[10]. Treat all of that as vendor-reported.

The engineering consequence is narrower than the marketing frame. Output-token throughput is one stage of observed latency; routing, prompt ingestion, retrieval, tool calls, safety checks, application orchestration and streaming behaviour all remain ^[15]. For a voice agent or a multi-step tool-using incident workflow, faster generation shrinks per-turn delay and leaves the rest of the budget untouched ^[15]. OpenAI says it uses the tier internally during incidents on logs, traces, team conversations, follow-up checks and preparing or validating fixes, with engineers retaining responsibility for judgment and deployment ^[16], and that some experiment loops which previously ran overnight now support several iterations in a workday ^[17].

Competitively, this is not unique ground: Anthropic already offers a fast mode for Claude ^[18]. On the supply side, mezha reports that Cerebras posted $193 million in quarterly revenue with a narrowed loss but guided to lower gross margin for the year, and its shares fell nearly 20 percent ^[19].

Watch three things. Whether pricing arrives as a per-token premium or a reserved-capacity contract, because that determines whether latency is a feature flag or a commitment. Whether workload-fit review persists as a rationing mechanism after general availability. And whether anyone publishes quality parity numbers for the Ultrafast serving path that are not produced by the two vendors selling it. Until then, build the fallback to Standard first and measure each latency stage separately.