OpenAI previewed a new processing tier called Ultrafast for its GPT-5.6 Sol model on August 13, 2026, running the model up to 14 times faster than standard processing and generating as many as 750 output tokens per second. The mode is powered by chipmaker Cerebras and is currently limited to a small group of customers through the OpenAI API, with wider access coming as capacity grows, according to OpenAI’s own announcement.
What Ultrafast Mode Actually Changes
Until now, getting near-instant responses out of a large language model usually meant trading down to a smaller, less capable model. OpenAI’s framing is that Ultrafast breaks that tradeoff: full GPT-5.6 Sol reasoning at a speed that used to require a lighter model. The 750 tokens-per-second figure puts it well past what most teams see from standard API calls, where response generation is often the visible bottleneck in a voice or chat interface.
It’s rolling out first through the API, not inside ChatGPT itself, and access right now runs through a waitlist rather than a flipped-on setting. OpenAI is asking interested businesses to submit their workload details before getting in.
Who Should Actually Care
This matters most for the use cases where a two- or three-second pause breaks the experience: voice assistants, live customer support, developer agents running multi-step tasks, and anything in financial or security response where a delayed answer has a real cost. A chatbot that types out a reply is one thing; a voice agent that goes silent for three seconds mid-conversation is a different problem entirely, and it’s the kind of gap our breakdown of what breaks first when you build an AI chatbot gets into.
For a business already running or planning an AI-driven support or agent workflow, this is worth watching even in preview. It’s a signal that speed-versus-capability is becoming a purchasing decision rather than a fixed limitation, and it’s one more argument for building AI features with the model layer decoupled from the rest of the product, so a faster tier can be swapped in later without a rebuild. That’s the kind of planning that comes up early in most of the AI development work we scope for clients.
What To Do About It Right Now
Preview access is limited and there’s no published pricing yet, so this isn’t something to build a roadmap around today. It is worth getting on the waitlist if latency is already a known pain point in a live product, and worth flagging internally if a team recently chose a smaller, faster model specifically to avoid GPT-5.6 Sol’s response time, since that calculation may not hold for long.
