For the last three years, the AI industry measured progress almost entirely in one direction: how smart a model could get. Bigger context windows, better reasoning scores, higher marks on coding benchmarks. Speed was an afterthought, something you traded away if you wanted the good answers. OpenAI’s newest release argues that era is ending.
On August 13, OpenAI began a limited preview of Ultrafast, a new processing tier for its GPT-5.6 Sol model that runs up to 14 times faster than the standard version, generating as many as 750 output tokens per second. The twist is what makes it possible. Ultrafast doesn’t run on Nvidia GPUs at all. It runs on wafer-scale chips built by Cerebras, a hardware startup that has spent years trying to convince the industry that GPUs are not the only way to make a large language model fast.
Quick facts: GPT-5.6 Sol Ultrafast
- Announced August 13, 2026, as a preview tier in the OpenAI API
- Up to 14x faster than standard GPT-5.6 Sol, topping out around 750 output tokens per second
- Runs on Cerebras wafer-scale engine hardware rather than Nvidia GPUs
- Limited to a select group of API customers during the preview, with capacity expanding gradually
- Already being tested in production for coding, commerce, financial research, and customer support
Why a chip nobody talks about is suddenly relevant
Nvidia’s GPUs dominate AI training and inference for a good reason: they are flexible, well supported, and available in the volumes hyperscalers need. But a GPU generates text one token at a time by repeatedly pulling model weights out of separate memory chips, across a physical distance, thousands of times per second. That trip between the processor and the memory is the single biggest bottleneck in how fast a language model can talk back to you.
Cerebras took a different approach. Instead of splitting a chip’s logic and its memory across separate pieces of silicon connected by wires, the company builds one enormous chip, roughly the size of a dinner plate, with the memory baked directly onto the same piece of silicon as the processing cores. There is no long trip for the data to make. According to Cerebras, that architecture is what lets GPT-5.6 Sol Ultrafast push out hundreds of tokens per second per user, a speed class that is normally reserved for much smaller, less capable models.
The tradeoff is capacity. Wafer-scale chips are expensive and difficult to manufacture at the volumes Nvidia ships GPUs in, which is exactly why OpenAI is starting Ultrafast as a small, invite-only preview rather than flipping it on for every developer at once.
How the 2026 speed race stacks up
Ultrafast lands in the middle of a year where nearly every major lab has been racing to make its models faster and cheaper at the same time, not just smarter. Here is how the current field compares on what’s publicly known.
| Model / Tier | Throughput | Hardware | Known for |
|---|---|---|---|
| GPT-5.6 Sol (Standard) | ~50 to 55 tok/s (implied) | Nvidia GPU clusters | Terminal-Bench 2.1 leader, ~89.5% per third-party trackers |
| GPT-5.6 Sol (Ultrafast) | Up to 750 tok/s (14x) | Cerebras wafer-scale engine | Real-time agents, coding, live support |
| Claude Opus 5 | Not disclosed | Mixed Nvidia / Trainium | 5M token context, ~89.1% Terminal-Bench 2.1 |
| Grok 4.6 | Not disclosed | xAI Colossus GPU cluster | Aggressive agentic coding push |
| MAI-Code-1.1-Flash | Faster than predecessor | Azure infrastructure | 75% cheaper coding tier |
Throughput figures as disclosed by each vendor or estimated from public statements, August 2026. Not all labs publish tokens per second.
Notice what is missing from most of that table: independently verified numbers. Speed claims in this industry are still mostly self reported, measured under best case conditions, and rarely tested by a neutral third party under identical load. Treat the 750 tokens per second figure the same way you’d treat any other vendor benchmark, as a ceiling under ideal conditions rather than what every request will get.
Why speed is becoming its own battleground
For most of the chatbot era, a slow answer was mildly annoying. In an agentic workflow, it’s disqualifying. An AI agent that books a table, debugs a pull request, or answers a customer support ticket has to make several model calls in a row, often waiting on the output of one before it can start the next. If each of those calls takes ten or fifteen seconds, the whole chain becomes too slow to use in anything resembling real time. Cut that latency by 14x and a workflow that used to take two minutes finishes in under ten seconds, which is the difference between a tool people tolerate and one they actually reach for.
That is also why this launch cannot be separated from the broader cost story playing out this year. Just days before Ultrafast appeared, OpenAI slashed its own API prices by 80 percent, only for DeepSeek to undercut it again within hours. Speed and price are turning into the same fight. A model that is both cheap and slow doesn’t get chosen for production agents, and a model that’s blazing fast but priced like a luxury good only gets used where latency truly justifies the cost. OpenAI appears to be betting it can eventually offer both, using Ultrafast as the proof that raw speed is achievable before figuring out how to make it affordable at scale.
The rest of the industry isn’t standing still
OpenAI is not the only lab treating speed as a selling point rather than a footnote. xAI has been pushing hard on the coding side with Grok 4.6, a model built specifically to handle long, autonomous coding sessions where a slow response breaks the developer’s flow just as badly as a wrong one. Meta has taken the opposite path to the same goal, optimizing for local speed instead of cloud speed. Its recently released Muse Glimmer agent model was built to run entirely on a single consumer GPU, using speculative decoding tricks to squeeze out several times the generation speed on hardware people already own, no data center required.
Put together, the pattern is clear. Every major lab has concluded that raw intelligence gains are hitting diminishing returns for a lot of everyday use cases, and that the next real differentiator is how quickly a model can turn a request into a finished action. Cerebras, previously a niche player known mostly to AI researchers, suddenly finds itself with a very public validation from the biggest name in the industry.
What to actually expect right now
If you’re a developer hoping to flip a switch and get 14x speed today, the honest answer is that you probably can’t yet. Ultrafast is in a narrow preview, available to a limited set of API customers that OpenAI selected directly, with access “expanding as capacity grows,” in the company’s own words. Cerebras hardware is not something OpenAI can spin up at Nvidia-scale volumes overnight, so anyone expecting broad general availability within weeks is likely to be disappointed.
Pricing for the Ultrafast tier also has not been published in detail, and given the specialized hardware involved, it would be reasonable to expect a premium over the standard tier, at least at first. The early customers testing it in production, across coding tools, commerce platforms, financial research systems, and support desks, are effectively the pilot group that will determine whether OpenAI expands the tier quickly or keeps it a boutique offering for high value, latency sensitive use cases.
The bottom line
Ultrafast is less about a single product than about what it signals. For the first time, one of the two or three most closely watched AI labs in the world has publicly staked part of its roadmap on non-Nvidia silicon, and framed raw response speed as being as important to the next generation of AI products as the underlying model’s intelligence. Whether Cerebras can supply that speed at a price and scale that makes sense outside a curated preview is the open question. But the direction is no longer in doubt. In 2026, being smart is table stakes. Being fast is starting to look like the actual competition.
