OpenAI and Cerebras Launch Ultrafast GPT-5.6 Sol Preview

OpenAI and Cerebras Systems have introduced a limited preview of Ultrafast mode, a high-speed service tier for the GPT-5.6 Sol model. The service utilizes Cerebras' Wafer-Scale Engine architecture to generate up to 750 output tokens per second, representing a speed increase of up to 14 times over standard processing.

The infrastructure achieves these speeds by utilizing 44 GB of SRAM per wafer to keep model weights on-chip, effectively removing memory-bandwidth bottlenecks. OpenAI designed the tier to support latency-sensitive applications, including financial research, security response, customer support, and voice interaction.

Performance benchmarks show that GPT-5.6 Sol Ultrafast completed the 2,500-question Humanity’s Last Exam in just over 11 hours. Access is currently restricted to a select group of customers while OpenAI evaluates the real-world utility of the service before expanding capacity.