OpenAI Announces a Model That Runs 14x Faster, Declines to Announce What It Costs
Ultrafast mode delivers the full GPT-5.6 Sol at up to 750 tokens per second on a chip the size of a dinner plate, and prices it in a unit the press release forgot to include.
On or around August 13, OpenAI previewed a new API tier called Ultrafast, which runs its GPT-5.6 Sol model at up to 14 times the speed of standard processing — up to 750 output tokens per second. The announcement was dense with numbers. It contained a 14, a 750, a 5.6, and a reference to a chip built from nearly an entire 300-millimeter silicon wafer. It contained, by careful accounting, exactly one number it did not want you to have, which was the price.
This is worth sitting with, because OpenAI has a well-documented pattern of charging for speed. Standard GPT-5.6 Sol runs $5 per million input tokens and $30 per million output tokens. The existing Fast mode, which is 2.5x quicker, roughly doubles that — $10 in, $60 out. Speed, in other words, is a paid multiplier, a lever you pull with a credit card, an escalator you ascend one dollar figure at a time: $5, then $10, then, for the tier that goes 14x, a serene and total silence.
The preview specifies that Ultrafast is available “initially to a select group of customers,” expanding “as capacity grows.” It does not say who those customers are, what they pay, or when anyone else may join them. It is a velvet rope in front of a door with no posted hours and no visible register. You cannot buy it, but you can be told, at 750 tokens per second, that you cannot buy it.
The genuinely impressive part — the part that is not a joke — is the engineering. Historically, “real-time” speed in AI meant picking a smaller model. You wanted fast, you took the dumb one. You traded quality for latency, and everyone understood the trade the way everyone understands that the express lane has a twelve-item limit.
Ultrafast’s pitch is that the trade is off. You get the complete GPT-5.6 Sol — not a distilled version, not a quantized understudy, not a smaller sibling wearing the big one’s badge — running at full frontier quality, just faster. It does this by running the model on Cerebras hardware, specifically the Wafer-Scale Engine, which is a single chip carved from almost an entire 300mm wafer. A normal chip is roughly the size of a postage stamp. This one is roughly the size of a dinner plate. The industry solved the speed-versus-intelligence tradeoff by declining to cut the wafer into individual chips, which is the kind of solution that is either a breakthrough or the setup to a joke about how they solved a problem by not doing a step.
It is, to be clear, a breakthrough. It is also a full-scale frontier model sprinting flat-out across a slab of silicon the size of something you’d eat off of, and there is no dignified way to describe that sentence.
OpenAI’s suggested use cases are where the deadpan writes itself. Ultrafast is pitched for incident response, customer service and support, financial market analysis, and e-commerce — the domains, in other words, where latency has a body count and a dollar figure. The unifying thread is emergencies, or at least situations urgent enough that waiting for tokens feels like watching a progress bar during a fire.
Which produces the tier’s actual value proposition, stated plainly: you can now be told something 14 times faster, in real time, precisely when it matters most. If the model is right, this is genuinely useful — a support agent that answers before the customer finishes typing, a market read that lands inside the window it describes. If the model is wrong, you now receive the wrong answer at 750 tokens per second, immediately, during your incident, with no perceptible pause in which anyone might have thought better of it. The wafer does not slow down for accuracy. The wafer does not slow down for anything. That is the entire point of the wafer.
Speed, notably, is agnostic about the correctness of what it is accelerating. This is true of all speed and has been since the invention of the second thing you regretted saying quickly.
Step back and the release is a small monument to where the industry has arrived. For three years the benchmark war was about how smart — reasoning scores, math olympiads, the model that could solve ten open problems and mention it in paragraph three. The frontier was a leaderboard of intelligence, and everyone raced up it.
Ultrafast is a flag planted on a different hill. The new axis is how fast, and the pitch is no longer “our model thinks harder” but “our model finishes sooner.” Tokens per second is becoming the number you put on the poster. It is a legible metric, it goes up and to the right, and — unlike intelligence, which is hard to measure and easy to argue about — everyone agrees on what a second is.
The one number still under embargo is the one that would let you decide whether any of this is worth it. OpenAI has told us the speed to three significant figures and the silicon to the width of the wafer. On the price it has offered a preview, a select group, and the promise of expansion as capacity grows — the corporate equivalent of a waiter reciting the specials, at length, warmly, and then declining to mention that they are market price.
You get the full model. You get it at 14x. You get it on a dinner plate. You will get the bill, presumably, faster than you can read it.