OpenAI previews GPT-5.6 Sol at 14x the speed

OpenAI previewed Ultrafast, a new API service tier, on August 13, running GPT-5.6 Sol up to 14 times faster than the standard tier at up to 750 output tokens per second. The speed comes from a partnership with Cerebras, whose wafer-scale chips power the mode; OpenAI says Ultrafast runs “the same intelligence” as GPT-5.6 Sol Standard, meaning no separate distilled or quantized model is doing the work.

OpenAI says a first batch of partner companies has been piloting Ultrafast across coding, commerce, financial-research, and support workloads, and it’s positioning the tier for live or near-production tasks: voice interfaces, customer support, developer agents, and security response, anywhere users are waiting on a response in real time rather than kicking off a background job.

What it means for you

Frontier-model latency has been the practical reason teams route voice and other real-time interfaces to smaller, faster models instead of a flagship one, and a 14x speedup that OpenAI claims doesn’t sacrifice intelligence directly attacks that trade-off. If you’ve been holding a frontier-quality feature back because response time made it unusable in a live conversation or voice call, this is worth a pilot before you assume you still need a smaller model for that use case, the same evaluation discipline behind treating any new speed or price claim as a number to verify yourself rather than take at face value, especially while it’s still limited to an early access group. It also puts pressure on other speed-focused releases this week, like Cursor’s faster cloud-agent builds, a sign the whole market is racing to close the latency gap between “frontier quality” and “usable in real time.”