OpenAI and Cerebras preview Ultrafast mode for GPT-5.6 Sol
OpenAI early on Aug. 14 previewed a new "Ultrafast Mode" for its flagship GPT-5.6 Sol model in partnership with AI chip company Cerebras. The new service tier raises output speed to as much as 750 tokens per second, up from roughly 53 tokens per second in the current Standard mode, a gain of as much as 14x with no quality reduction, according to the companies. OpenAI said the mode will launch first through its API and is already in limited preview for selected customers. Cerebras also shared benchmark results showing GPT-5.6 Sol Ultrafast completing Humanity's Last Exam, a 2,500-question benchmark, in 11 hours and 11 minutes. Claude Fable 5 took 78 hours and 27 minutes on the same test, with the companies saying accuracy was similar while Sol ran at nearly seven times the speed. OpenAI highlighted use cases in legal drafting, financial modeling, engineering reports, incident response, research, customer support and commerce. The speedup is tied to Cerebras' wafer-scale hardware design, including its WSE-3 chip and a pipelined multi-wafer setup intended to reduce memory bandwidth bottlenecks during large-model inference.








