Standard GPT-5.6 Sol types out roughly 53 tokens per second — already faster than most people talk. OpenAI just previewed a mode that hits 750. That's not a faster typist, that's a different category of interaction entirely, and it didn't get there by making the model smarter. It got there by borrowing someone else's chips.

14x the Speed, Same Brain, Borrowed Silicon

OpenAI's new Ultrafast service tier runs GPT-5.6 Sol on Cerebras wafer-scale hardware instead of its usual infrastructure, pushing output up to 14 times faster than Standard mode. Crucially, OpenAI is clear that the model's intelligence, context window, and answer quality are unchanged — only the speed at which tokens arrive is different. It's currently in limited preview with select customers, including Jane Street, Podium, Basis, and Rogo, with no public pricing or general availability date yet. Standard GPT-5.6 Sol currently runs $5 per million input tokens and $30 per million output tokens, and Ultrafast's premium over that, if any, hasn't been disclosed.

Speed Is the Feature Nobody Benchmarks

Everyone obsesses over model intelligence leaderboards, but latency is the thing that actually determines whether an AI product feels alive or feels like a chatbot from 2019 with better vocabulary. At 750 tokens per second, voice agents stop having that awkward "thinking..." pause, live coding assistants stop feeling like autocomplete with a lag, and real-time trading or customer-service tools become genuinely viable instead of "technically possible but painful."

The bigger story here might be the Cerebras partnership itself: OpenAI reaching outside its own infrastructure for a specific job is a tacit admission that no single hardware stack wins every use case. Expect more of this kind of hardware-shopping as inference speed becomes as competitive a battleground as raw model capability.

750 tokens a second means the AI isn't just answering your question anymore — it's finishing your sentence before you've decided how to end it.

When response speed becomes part of the user experience, integrating it well is the whole game — that's the kind of AI implementation work WTK does for businesses, so let's talk about building yours to actually feel fast.

Source: OpenAI