OpenAI Previews Ultrafast Mode: GPT-5.6 Sol Hits 750 Tokens Per Second
A new API tier powered by Cerebras silicon brings 14x faster inference speeds to OpenAI's flagship models, eliminating the latency tax on multi-step agents.
If you build multi-step AI agents, you know the waiting game all too well. An agent plans, calls a tool, checks the output, reflects, and formats a response. Even if each step is moderately fast, three or four back-to-back LLM calls can easily turn a snappy user experience into a fifteen-second loading spinner. OpenAI's new 'Ultrafast' mode is aimed squarely at killing that latency.
The new preview tier runs GPT-5.6 Sol on Cerebras wafer-scale hardware, pushing output speeds up to 750 tokens per second. That is roughly 14 times faster than standard cloud inference runs today. Instead of streaming text at reading speed, the model can generate entire pages of code or complex structured reasoning traces in the blink of an eye.
Why Speed Changes Agent Architecture
Raw generation speed is not just about making chatbots feel snappier. When inference happens this fast, how you design agent workflows changes fundamentally:
- Denser verification loops: You can have a secondary model review, lint, or run unit tests on generated code multiple times without blowing your latency budget.
- Real-time voice and interactive agents: Sub-second reasoning loops mean conversational agents can think before speaking without awkward pauses.
- Tree-of-thought exploration: Agents can sample several different branches of problem-solving simultaneously and pick the best path before taking action.
What This Means for Builders
Keep an eye on token economics as these high-speed endpoints roll out broadly. While ultrafast inference lets you run much more ambitious autonomous loops, it also makes it easier to burn through tokens if your exit conditions and error-handling guardrails are loose. Fast inference demands tighter architecture, not sloppy prompting.