Cerebras powers OpenAI's GPT-5.6 Sol inference engine, delivering 750 tokens/second and processing 2,500 questions in 11 hours

Cerebras Systems announced the launch of Ultrafast mode, a new service tier for OpenAI's GPT-5.6 Sol that enables processing speeds of up to 750 tokens per second—up to 14 times faster than standard processing. Initially available in limited preview to select OpenAI customers, Ultrafast provides access to the full intelligence of GPT-5.6 Sol with significantly reduced latency.

According to CEO Andrew Feldman, achieving such speeds is crucial for wider AI adoption. "Every major computing shift has been unlocked by a leap in speed," Feldman said. "The PC era needed the jump from kilohertz to gigahertz, and the internet needed the move from dial-up to broadband before it could reach everyone. AI is no different. Now that frontier models are demonstrably capable, the next constraint on adoption is how fast they run."

Sachin Katti, VP Compute Strategy & GPT-Infra at OpenAI, noted that the partnership is exploring how customers benefit when they can access the intelligence of frontier models with significantly lower latency. "We're starting with a small group of customers to learn where that speed creates meaningful value, and we'll use those learnings to inform how we expand the service over time," he said.

The Ultrafast tier addresses a fundamental constraint: organizations have historically had to choose between the capabilities of larger models and the faster response times of smaller models. Based on output speeds reported by Artificial Analysis, Ultrafast is 5x faster than Claude Opus 4.8 in Fast mode and 11x faster than Claude Fable 5, combining frontier intelligence with unprecedented speed.

Benchmark results demonstrate the practical impact. On Humanity's Last Exam—a 2,500-question benchmark spanning graduate-level chemistry, economics, and literature—GPT-5.6 Sol Ultrafast answered the full question set in just over 11 hours, compared to more than three days of continuous compute for Claude Fable 5, reaching comparable accuracy nearly 7x faster. On GDP-Val, a benchmark of economically valuable knowledge-work tasks including legal briefs, financial models, and engineering reports, Ultrafast delivered a 5.6x end-to-end speedup with no loss in quality.

Ultrafast's speed derives from Cerebras' Wafer-Scale Engine architecture, which keeps model weights on-chip—44 GB of SRAM on each wafer-sized chip—rather than shuttling them between on-chip and off-chip storage as GPU-based inference requires. This approach eliminates the memory-bandwidth bottleneck that constrains frontier-model inference speed on conventional hardware. Cerebras plans to expand Ultrafast capacity to additional customers based on initial feedback and experiences.