OpenAI Previews Ultrafast Mode: What 14x Speed Changes for Voice and Checkout

August 14, 2026. OpenAI announced Ultrafast on August 13, a new API service tier that runs GPT-5.6 Sol up to 14 times faster than Standard processing, generating up to 750 output tokens per second on hardware from Cerebras. Two of the five scenarios OpenAI puts on the announcement page are the two that decide revenue for a small business: a voice agent that resolves a complex call without stalling, and a store assistant that answers a product question while the shopper is still deciding. So this is not a benchmark story. The practical question for anyone running a voice agent, a booking bot or a checkout assistant this quarter is narrower and more awkward: there is no published price, no date, and the one named customer whose buyers look like yours is a software vendor rather than a business. Here is what the announcement supports, and what it does not.

What OpenAI actually announced

1. A new service tier, not a new model. Ultrafast runs the existing GPT-5.6 Sol. OpenAI describes it as a new speed class for frontier intelligence, launching first in the OpenAI API.
2. Up to 14 times faster than Standard, and up to 750 output tokens per second. Those are the only two speed figures on the page.
3. Powered by Cerebras. OpenAI calls it the next step in an existing partnership to bring low latency inference to its platform.
4. Limited preview. Available to a select group of customers, with access expanding as capacity grows. There is a signup form for updates, and no general availability date.
5. Four named early customers. Jane Street, Podium, Basis and Rogo, testing across coding, commerce, financial research and support.

The claim underneath all of it is the one worth holding onto. OpenAI writes that until now, getting real time speed typically meant choosing a smaller or more specialized model. If that trade is genuinely gone, the design rule most teams have been following since 2024, use the small model for the live turn and the big model for the background job, stops being a law and becomes a budget decision. TechCrunch reported the launch the same day and noted that Anthropic already ships a fast mode of its own, so the tier itself is becoming a normal part of how frontier models are sold.

The number published, and the one that is not

750 output tokens per second is a throughput figure. It tells you how fast text comes out once text is coming out. For the two use cases OpenAI itself puts in front of small businesses, voice and checkout, the number that decides whether the thing feels instant is a different one: how long the caller sits in silence, or the shopper watches a blinking cursor, before the first word arrives.

OpenAI published the multiple and the throughput. It did not publish a time to first token figure for Ultrafast. That is not a criticism of the release, it is a caution about how the release will be quoted. A voice agent that begins speaking in 400 milliseconds and then speaks at a normal pace feels better on a phone call than one that pauses for two seconds and then delivers the answer in a burst. Both can honestly report a large speedup in tokens per second. Podium is the customer that speaks to this directly. Its product lead for voice AI is quoted saying the speed changes the call experience for the more complex work, which is a statement about calls that involve several steps, not about every call.

So the practical reading is that Ultrafast should help most where the model was doing several things before answering. If your agent answers simple questions from a short prompt, the model was probably never the slow part.

A service tier with no published price

The announcement links a signup form, not a pricing page. Compare that with how OpenAI handled the adjacent tier nine days earlier: the August 5 changelog entry extending Fast mode to long context requests over 272K tokens quotes speeds up to 2.5 times faster than Standard and points straight at pricing details. Ultrafast does not.

For an agency, that difference is the whole story this week. You cannot put a number in a proposal for something whose unit cost is unpublished and whose access is gated. Anyone quoting a build on Ultrafast in the next month is quoting a variable they do not control. The honest version of the sentence to a client is that a faster tier now exists, it is not generally available, and when it is priced we will tell you whether it changes the recommendation.

The customer list tells you who this is for

Look at who OpenAI named. Jane Street is a quantitative trading firm. Basis and Rogo build software for finance and accounting teams. Those three are enterprises or vendors selling into enterprises. Podium is the fourth, and Podium sells messaging and lead response software to local businesses: home services, auto, healthcare, the exact buyer profile a small agency serves.

Notice what that means. The one company on the list whose customers resemble a plumbing firm or a dental practice is a vendor, not the plumbing firm. Small businesses will not receive Ultrafast by making an API decision. They will receive it, if they receive it, as a quiet improvement inside a product they already pay for. That is a genuinely useful conclusion, because it tells an operator where to direct the question. Ask your voice or CRM vendor what inference tier they run and whether they intend to move. That is a five minute email and it is worth more than a rebuild.

Capacity belongs to Cerebras, so there is no date

OpenAI says it will expand access as capacity grows. Read that literally. Ultrafast runs on Cerebras hardware, so the constraint on how fast the preview opens up is a partner's deployed fleet, not a rollout calendar OpenAI controls alone. There is no announced general availability date anywhere in the release.

Which produces one concrete rule. Do not place Ultrafast on a client roadmap with a quarter attached to it. A dependency with no price and no date is a research item, not a milestone. Put it in the section of the plan where you list things you are watching, and keep the current build working on a tier you can actually buy today.

Speed just got cheaper at the fastest layer

OpenAI's own commerce example deserves a second read. It describes answering product questions, checking inventory, personalizing recommendations and resolving checkout issues while the shopper is still deciding, before hesitation becomes an abandoned cart. Every one of those is a latency problem. But in a real store, walk the round trip and ask which piece is slow. The inventory lookup against an ERP. The shipping rate call. The app that adds 900 milliseconds to every page. The order API that rate limits you at checkout. In most small and mid sized stacks, the language model was already among the faster components, and it is the one that just got 14 times faster.

This is the part almost nobody will write, and it is not a knock on the release. It is an instruction. Before you buy speed, measure where your seconds are actually going. If the model answers in 900 milliseconds and your inventory call takes 2.4 seconds, a 14 times faster model moves your total from roughly 3.3 seconds to about 2.5, and the shopper still waits. The same arithmetic applies to a voice agent that has to query a calendar before it can offer a slot. Speed you can buy is only worth buying at the layer that is slow, and finding that layer costs you an afternoon with a log and a stopwatch, not a contract.

What it means for operators

Nothing about this release requires action this week from a business that is not already running a live AI conversation with customers. For those that are, there are four moves, in order of value.

1. Instrument before you optimize. Log the elapsed time of each external call in your agent's turn, separately from model time, for one week. You will usually find the bottleneck is not the model, and that finding is worth more than any tier.
2. Ask your vendor, do not rebuild. If you run a voice agent or chat widget through a platform, send one email asking which inference tier it uses and whether it plans to adopt faster tiers. Their answer determines whether this release reaches you at all.
3. Separate the two speed questions in your own build. Time to first response and total response time are different problems with different fixes. Streaming, a short acknowledgement, and a smaller model for the opening turn all address the first. Only raw throughput addresses the second.
4. Keep the tier out of the contract. Quote what you can buy today. Revisit when a price exists.

We build and maintain this layer for clients every week, from AI voice agents that have to answer inside a live call to AI automation that connects the model to the systems holding the slow data. If you want an honest read on where your own seconds are going before you spend anything on speed, that is the audit to run first, and it is the same one our AI engineers start with on a new build.

Two limits on everything above. Ultrafast is in limited preview, so the performance any given team sees is not yet a general result. And all figures here come from OpenAI's announcement page dated August 13, 2026, which may be updated as the preview expands.