OpenAI's Ultrafast Mode Pushes GPT-5.6 Sol to 750 Tokens a Second

OpenAI is now running GPT-5.6 Sol at 14 times its normal processing speed through a new mode called Ultrafast, announced on August 13, 2026. The feature pushes output to as much as 750 tokens per second, a pace that turns tasks once measured in hours into ones measured in minutes.

The gain comes from hardware, not just software. Ultrafast runs on Cerebras' Wafer-Scale Engine, a chip architecture built around 44 GB of integrated SRAM memory that keeps data close to compute rather than shuttling it across a network of GPUs.

That architecture explains the scale of the speedup better than any prompt-engineering trick could. Cerebras has spent years positioning wafer-scale silicon as an inference accelerator, and OpenAI's adoption of it for a flagship model is a concrete bet that the approach can now serve production traffic.

The benchmark numbers make the case vivid. Running the 2,500-question "Humanity's Last Exam" test, Sol Ultrafast finished in 11 hours and 11 minutes, compared with 78 hours and 27 minutes for OpenAI's own Fable 5.

Independent tracking from Artificial Analysis puts Sol Ultrafast at roughly 11 times faster than Fable 5 and about 5 times faster than Anthropic's Opus 4.8 running in its own Fast mode. That comparison places OpenAI's latency advantage squarely against its closest commercial rival rather than against its own prior generation alone.

Where the speed matters

OpenAI is pitching Ultrafast for jobs where waiting costs money or safety margin: incident response, financial market analysis, customer service, e-commerce, voice applications, coding, and research workflows. In each case, the value isn't smarter answers, it's faster ones, delivered while the underlying situation is still current.

For a trading desk or an on-call engineer, a model that replies in fractions of a second changes what's practically usable in the loop, not just what's theoretically possible.

Access remains limited for now. Ultrafast is in preview and rolled out to a small group of customers, with OpenAI saying it will widen availability as Cerebras capacity grows.

That caveat matters for anyone evaluating the mode today. The headline speed figures describe what the system can do under current conditions, not what every developer can expect to reserve immediately, and broader rollout will depend on how quickly wafer-scale inference capacity can be added.