What's this about?
OpenAI previewed Ultrafast, a new API service tier that runs its top model, GPT-5.6 Sol, up to 14 times faster than normal — up to 750 output tokens per second — with no reduction in quality. It's powered by Cerebras hardware rather than conventional GPUs. The headline demonstration: Sol on Ultrafast answered all 2,500 questions of Humanity's Last Exam in 11 hours and 11 minutes, where Claude Fable 5 needed 78 hours and 27 minutes. It's a limited preview with no announced pricing.
What exactly is Cerebras?
Almost all AI runs on GPUs — chips originally built for video games, wired together in large numbers. Cerebras took a different path: it builds a single chip the size of an entire silicon wafer, roughly a dinner plate. The reason this matters for speed is that generating text is sequential — each word depends on the last — so the limiting factor isn't raw computation but how fast data moves between chips and memory. Keeping the whole model on one enormous chip removes most of that movement. The trade-off is that these machines are expensive and scarce, which is likely why the preview is invitation-only with no published price.
What's the community saying?
- Speed is underrated, and this thread said so loudly. The most-supported argument was that latency changes what you can build, not just how pleasant it feels — below a certain threshold an agent can participate in interactive work instead of being a batch job you wait on.
- The absent price tag was immediately noticed. With no pricing published, commenters concluded it's either "if you have to ask" territory or OpenAI gauging demand before committing.
- It rearranges the speed/intelligence frontier. Several observed that Gemini 3.7 Flash, released hours earlier, no longer sits on the Pareto frontier of speed versus intelligence — the cheap-and-fast tier just lost its main justification against a top model at 750 tokens/sec.
- The Cerebras partnership finally paid off. Long-time watchers noted they'd been waiting years for something impressive from the OpenAI–Cerebras collaboration, and the Humanity's Last Exam timing comparison was the number that convinced them.
- Hardware speculation. A popular thread compared this to dedicated video encode/decode silicon, predicting eventual thumbnail-sized user-replaceable modules with a model baked into the hardware.