OpenAI Launches Ultrafast Mode: GPT-5.6 Sol Runs 14× Faster, Powered by Cerebras

At the heart of Ultrafast mode is a tight integration between hardware and software. Traditional large language model inference is usually constrained by GPU memory bandwidth and compute-unit utilization — especially when generating long text, the serial dependency between tokens causes noticeable response latency. By partnering with Cerebras, OpenAI deploys GPT-5.6 Sol on Cerebras' Wafer-Scale Engine. This chip treats an entire wafer as one massive compute unit, avoiding frequent data movement between multiple chips and dramatically reducing memory latency. The result is a model that can sustain 750 output tokens per second during generation, whereas standard modes typically deliver only a few dozen tokens per second. For users, this means answers that once took several seconds — or even tens of seconds — can now be produced almost instantly, moving the experience from character-by-character output into true synchronous conversation. For latency-sensitive applications such as chatbots, code assistants, real-time translation, and voice assistants, this upgrade represents a qualitative leap in user experience rather than a mere change in numbers.

Two benchmark results published by Cerebras provide strong evidence for Ultrafast mode. The first is Humanity's Last Exam (HLE), a notoriously difficult evaluation set containing 2,500 questions spanning mathematics, science, and the humanities. GPT-5.6 Sol Ultrafast answered all questions in 11h11m, while Anthropic's Claude Fable 5 needed 78h27m for the same task — a gap of roughly 7×. The second is the GDP-Val benchmark, which measures end-to-end speed; it showed a 5.6× end-to-end speedup with no quality degradation whatsoever. Notably, speed and quality have historically been seen as a trade-off — acceleration usually means trimming computation and sacrificing accuracy. This result suggests that with better hardware architecture, both can be achieved at once.

Rohan Varma from OpenAI's product team described the launch this way: "With GPT-5.6 Sol Ultrafast, Cerebras enables AI that keeps up with how you think, code, and collaborate." The quote captures the real value of Ultrafast mode — not just a numbers game, but bringing AI's response speed close to the speed of human thought, thereby fundamentally changing how people interact with machines. In terms of concrete applications, OpenAI highlighted three categories of tasks: legal briefs, financial models, and engineering reports. What these scenarios share is that the documents are long, structurally complex, and require repeated iteration, making them extremely sensitive to generation speed. In the past, producing a dozens-of-pages engineering report meant waiting again and again; now, real-time generation lets lawyers, analysts, and engineers collaborate with AI as naturally as they would with a colleague.

The impact of this launch on the broader AI industry should not be underestimated. For a long time, the AI race focused primarily on model capability — whose model scores higher on benchmarks and can answer more complex questions. With Ultrafast mode, OpenAI extends the battlefield from “what can it do” to “how fast can it do it.” Faster inference is not only about user experience but is also directly tied to cost — processing more requests in the same amount of time means a significant drop in unit compute cost, which matters especially for enterprises that call APIs at scale. Moreover, as a chip company focused on inference acceleration, Cerebras' partnership with OpenAI signals that specialized inference chips are moving from the fringe into the mainstream, competing head-to-head with NVIDIA's general-purpose GPU approach. It is foreseeable that in the coming months, competitors such as Anthropic and Google will likely roll out similar acceleration offerings in response — a new arms race around inference speed has already begun. At the same time, faster inference also means the same hardware can serve more users per unit of time, spreading out the marginal cost of each inference, which may push API prices further down and let more developers and small teams afford high-performance AI services.