GPT-5.6 Sol Ultrafast hits 750 tokens/s

Ultrafast makes inference latency a product feature rather than an infrastructure footnote, especially for agentic workflows that generate large volumes of tokens. The tradeoff will be cost and limited availability, but the pairing demonstrates why specialized inference hardware is becoming strategically important.

• –750 tokens per second can materially shorten coding-agent loops and interactive applications
• –Cerebras’ wafer-scale systems give OpenAI a differentiated speed tier without changing GPT-5.6 Sol’s underlying capabilities
• –The initial select-customer rollout suggests capacity, economics, and reliability remain constraints
• –Developers should benchmark end-to-end latency, not just output speed, including queueing, tool calls, and time to first token
• –Faster reasoning makes higher-effort model modes more practical, but could also accelerate token consumption and API spend