OpenAI and Cerebras unveiled Ultrafast on August 13 — a new API tier that runs GPT-5.6 Sol at up to 750 output tokens per second, which OpenAI describes as up to 14 times its Standard tier's speed. The model's intelligence, context window, and output quality don't change; only how fast the tokens arrive does. It's OpenAI's third GPT-5.6 Sol speed option: Standard, then the existing Fast mode at roughly 2.5 times Standard's speed for about 2 times the price, and now Ultrafast on top of both. Access is limited to a small group of early customers for now, spanning coding, financial research, voice AI, and e-commerce, with OpenAI saying it wants to study where the speed creates real value before expanding further.
The speed comes from Cerebras's wafer-scale chips rather than GPUs: each wafer-sized chip carries 44GB of on-chip SRAM, so a model's weights sit directly on the chip instead of shuttling back and forth from separate memory — the round trip that caps how fast a GPU cluster can serve tokens per second. Cerebras's own announcement reports an 11x speedup over Anthropic's Claude Fable 5 and a 5.6x end-to-end gain on the GDPVal benchmark with no measured quality loss.
What's disclosed about GPT-5.6 Sol's speed tiers — and what isn't
- 2.5x · Fast mode
- OpenAI's existing faster tier, already public.
Includes: Roughly double Standard's price for 2.5x the speed, per independent reporting.
Excludes: An OpenAI-published multiplier — the 2x figure is outside reporting's characterization, not a rate OpenAI states directly. - 750 tok/s · Ultrafast (new)
- Up to 14x Standard's speed, per OpenAI and Cerebras's joint announcement.
Includes: Confirmed benchmark gains (11x vs. Fable 5, 5.6x on GDPVal) from Cerebras's own writeup.
Excludes: Any published price, and any confirmed general-availability date — limited preview only.
The timing lands one day after Google shipped Gemini 3.7 Flash, a launch that put inference speed back at the center of frontier-model marketing. OpenAI hasn't framed Ultrafast as a response to that release, but The Decoder's own read is blunter about the business logic either way: running a top model at 14x speed on rented specialist hardware is unlikely to be cheap, and a tiered-speed pricing ladder gives OpenAI a direct cut of whatever revenue gain the extra latency savings create for a customer.
- OpenAI and Cerebras previewed Ultrafast, a third GPT-5.6 Sol speed tier, on August 13, 2026.
- It runs at up to 750 output tokens per second — up to 14x OpenAI's Standard API tier — with the model itself unchanged.
- Cerebras's wafer-scale chips keep the model's weights on-chip, avoiding the memory transfers that cap GPU speed.
- Access is limited to a small group of early customers across coding, finance, voice, and e-commerce.
- Caveat: no price and no general-availability date have been published for either company's own announcement.