OpenAI Puts Latency on the Price List

OpenAI opened a limited preview of Ultrafast on August 13th, an API service tier that the company says runs its flagship GPT-5.6 Sol at up to 750 output tokens per second on Cerebras hardware, up to 14 times the speed of its Standard processing tier ^[1][2][3]. The consequence is not the headline number: it is that response speed is now a separate product line rather than something you obtained by downgrading to a smaller model ^[4].

Run the 14x claim backwards and the implied Standard-tier rate is roughly 54 output tokens per second ^[5]. On a 1,500-token answer, that is about two seconds against about twenty-eight ^[6]. Twenty-six seconds is the gap between a loop a human stays inside and a loop a human abandons, and it is the reason several application shapes have been off the table. OpenAI is pointing the tier at exactly those shapes: incident response, voice agents, financial research, customer support, commerce and interactive scientific work ^[7]. Its own framing is that developers can redesign an application around lower latency, while a faster chat response mostly changes how an existing interface feels ^[8].

The named early testers are Jane Street, Podium, Basis and Rogo ^[9]. Podium product lead Courtland Lykins said Ultrafast had been valuable in the company's voice stack because "the speed completely changes the call experience" for complex work ^[10]. Rogo told OpenAI the tier made complex analysis feel closer to real-time interaction ^[11]. Internally, OpenAI says engineers are using it to read logs, traces and discussions during outages, and researchers are trying to compress experiment cycles that previously ran overnight into repeated same-day iterations ^[12][13].

Now the caveats, which matter more than the testimonials. The 750-token figure measures output generation and does not by itself establish time to first token or the wall-clock cost of a workflow that calls tools and external systems ^[14]. The 14x comparison is OpenAI's own measurement against its own Standard tier ^[15]. Real production numbers will move with prompts, reasoning settings, output length and the surrounding application ^[16]. And the schedule has already stretched: OpenAI disclosed the 750-token service in its June 26th GPT-5.6 Sol preview and said it would launch in July, the model reached general availability on July 9th, and the Cerebras-backed tier entered limited preview on August 13th, 35 days later ^[17][18][19]. Access remains restricted to selected customers, widening as compute capacity grows ^[20].

The supply side is a January 14th agreement to add 750 megawatts of low-latency capacity in stages through 2028, from a December 2025 master relationship agreement that also gave OpenAI an option on another 1.25 gigawatts by the end of 2030 ^[21][22]. Cerebras attributes the throughput to its wafer-scale design, which keeps model weights in on-chip memory ^[23][24].

OpenAI is also on both sides of the trade. In July it exercised warrants for 10,033,508 nonvoting Class N shares at $0.00001 each, a cash cost of about $100.34, for 4.22% of Cerebras ^[25][26][27]. Cerebras disclosed it on page 29 of a quarterly filing after the close on August 12th; at the roughly $229 Class A price the stake implied about $2.3 billion ^[28][29]. Warrants for 23,411,518 more shares remain, taking OpenAI to about 12.8% if all vest ^[30][31]. Cerebras shares fell on August 13th anyway ^[32].

Watch for a published time-to-first-token figure, for general availability rather than allowlist, and for what the tier costs relative to Standard. Until then, treat Ultrafast as a capacity experiment you can prototype against, not a latency budget you can promise a customer.