Cerebras Accelerates GPT-5.6 Sol for Ultrafast Inference

Cerebras has released benchmark results demonstrating the performance of OpenAI’s GPT-5.6 Sol model operating on its Ultrafast inference architecture. The evaluation utilized Humanity’s Last Exam, a rigorous assessment comprising 2,500 questions designed to require doctoral-level expertise across specialized disciplines such as chemistry, economics, and literature. In the direct comparison, GPT-5.6 Sol deployed on Ultrafast hardware processed the entire benchmark in 11 hours and 11 minutes. By contrast, the same examination required 78 hours and 27 minutes of continuous compute when evaluated using Claude Fable 5 under standard conditions. This performance differential represents a near sevenfold increase in processing speed, enabling the model to traverse complex academic material within a single operational day. Cerebras reported that Ultrafast maintained accuracy metrics comparable to conventional inference methods, underscoring the architecture’s capacity to dramatically reduce latency without compromising output quality. The results highlight a significant advancement in high-throughput AI evaluation, suggesting that ultrafast hardware implementations could streamline research validation and accelerate deployment cycles for frontier language models.