OpenAI just put speed on the price list as a product of its own
On 13th Aug’26, OpenAI previewed Ultrafast, an API service tier that runs GPT-5.6 Sol at up to 750 output tokens per second. That’s around 14x standard processing. It runs on Cerebras wafer-scale hardware rather than the GPU fleet behind everything else, off the back of a partnership announced in January. It is a limited preview to selected customers, and there is no release date. It becomes the third rung of a ladder: Standard, then Fast, that’s 2.5x the speed at about double the price. Now Ultrafast above both. OpenAI says its own engineers use it to read logs during a live outage, and that overnight research batches have turned into same-day loops.
What it actually means
Read this next to the DeepSeek story and the week makes sense. One provider is now charging by when you use it. The other is charging by how fast you want it back. Neither is charging purely for intelligence any more. Compute is being priced like a utility by looking at scarcity, timing, grade. The useful read for builders isn't 'buy the fast tier.' It's that latency, not capability, has been the quiet killer of a lot of good ideas. Write down which of your shelved ideas died of slowness rather than of quality. That list is about to be worth something so keep the realism on! The one number that matters here hasn't been published yet, and a tier that costs four times standard changes the maths on every one of those ideas. Don't rebuild anything this week.
Latency was never a law of nature.
750/secoutput tokens per second, around 14x standard
Source: OpenAI · Saturday, 15 August 2026