OpenAI just made its frontier model run 14 times faster than its standard mode, and they did it by plugging into someone else's hardware.
The Summary
- OpenAI's new Ultrafast API tier runs GPT-5.6 Sol at up to 14X normal speed, delivering up to 750 output tokens per second
- The speed boost comes from Cerebras hardware, marking OpenAI's first major public integration with a third-party inference provider
- OpenAI is simultaneously releasing GPT-5.6 with smarter model selection and a new Responses API designed to help builders choose the right model for each task
- The move signals OpenAI prioritizing developer experience and cost efficiency over keeping everything in-house
The Signal
OpenAI is doing something it rarely does: admitting it needs help. The Ultrafast tier runs on Cerebras chips, not OpenAI's own infrastructure. Cerebras makes wafer-scale processors, chips the size of dinner plates that excel at parallel processing. For inference speed, they're apparently faster than anything OpenAI has in its own data centers. The 750 tokens per second throughput means GPT-5.6 Sol can generate roughly 500 words in a single second.
This matters because speed is the new battleground for AI agents. An agent that takes 30 seconds to respond loses to one that responds in 2 seconds, even if the slower one is slightly smarter. OpenAI's builder's guide explicitly frames GPT-5.6 around helping startups build "faster, more cost-efficient AI agents", which tells you where they think the market is heading. Not chatbots. Not content generation. Agents.
"Speed is the new battleground for AI agents. An agent that takes 30 seconds to respond loses to one that responds in 2 seconds."
The new Responses API is the other piece of this. It's designed to help developers automatically route tasks to the right model, which means OpenAI is acknowledging what builders already know: you don't need GPT-5.6 for every task. Use the big model for reasoning. Use the small model for formatting. Use Ultrafast when latency kills your product. The API handles the switching.
What's notable is the timing. OpenAI is releasing a speed tier and a model selection framework at the same time. That's not a coincidence. They're building the infrastructure for a world where AI applications run dozens or hundreds of model calls per user interaction. If you're building an agent that researches, plans, executes, and reports back, you might hit the API 50 times in one workflow. Ultrafast makes that viable.
The Cerebras partnership also tells you something about OpenAI's production constraints:
- They don't have enough fast inference capacity in-house
- They're willing to route premium traffic to third-party hardware
- Speed matters more than keeping the full stack proprietary
Cerebras gets legitimacy. OpenAI gets speed without building new data centers. Developers get sub-second response times for the first time on a frontier model.
The Implication
If you're building on OpenAI's API, test Ultrafast immediately. Any application where users wait for responses, where agents need to chain multiple calls, or where real-time interaction matters, just got 14X more viable. The latency wall that made certain agent architectures impractical just dropped.
Watch for other model providers to announce their own speed tiers in the next 60 days. Anthropic and Google won't let OpenAI own "fastest frontier model" without a fight. And watch Cerebras, they just became the inference hardware company that proved it could handle OpenAI-scale traffic.