GPT-5.6 Sol Ultrafast Is 14x Faster Without Cutting Model Power

GPT-5.6 Sol Ultrafast mode is officially here, giving OpenAI customers access to the same GPT-5.6 Sol model at speeds reaching 750 output tokens per second.

OpenAI announced GPT-5.6 Sol Ultrafast as a new high-speed processing option that can run up to 14 times faster than GPT-5.6 Sol’s Standard mode.

Unlike approaches that increase speed by switching to a smaller model, Ultrafast continues to use the same GPT-5.6 Sol model.

Cerebras hardware pushes Sol to 750 tokens per second

Ultrafast runs on Cerebras’ Wafer-Scale Engine architecture, which includes 44GB of on-chip SRAM.

Keeping more data directly on the chip reduces the data movement bottlenecks that can slow AI inference.

OpenAI first revealed plans for Cerebras-powered Sol inference in June 2026, initially targeting a July launch. The company later offered the system through a private preview with selected customers.

GPT-5.6 Sol Ultrafast outpaces rival models

Performance figures from Artificial Analysis highlighted by Cerebras show a large speed advantage.

GPT-5.6 Sol Ultrafast runs around 11 times faster than Fable 5 and about five times faster than Opus 4.8 running in Fast mode.

The difference also appeared during Humanity’s Last Exam. Sol Ultrafast processed all 2,500 questions in 11 hours and 11 minutes, while Fable 5 required 78 hours and 27 minutes. Sol delivered comparable accuracy.

Access remains limited for now

OpenAI says Ultrafast could suit workloads where response speed matters, including incident response, financial research, real-time customer support, and research workflows.

Access remains limited because Cerebras capacity is currently constrained. OpenAI will evaluate customers based on workload fit and available capacity.

The launch follows another recent expansion of OpenAI’s GPT-5.6 lineup, with the company also releasing GPT-5.6-Cyber.