It finished in 11 hours what used to take 78

Here's the deal: on August 13, OpenAI previewed Ultrafast mode with Cerebras — a new service tier for GPT-5.6 Sol. The claim fits in one sentence. Same model, up to 14x faster, no change in quality.

The numbers: output runs at up to 750 tokens per second, as much as 14x standard processing. And this is not a shrunken or distilled variant — it's the full GPT-5.6 Sol. That distinction matters. Speed improvements usually mean dropping to a smaller model or lowering reasoning effort. This one leaves the model alone and swaps the hardware underneath.

The most striking demonstration is benchmark wall-clock. GPT-5.6 Sol Ultrafast completed 2,500 questions of Humanity's Last Exam in 11 hours 11 minutes. Claude Fable 5 took 78 hours 27 minutes on the same set — a 7x gap. On GDP-Val, the setup posted a 5.6x end-to-end speedup with no quality loss.

This is the moment AI competition grew a third axis. Intelligence (benchmark scores) and price were the two. Speed just joined them.

What each company brought

Cerebras builds chips out of entire wafers. Standard semiconductor manufacturing dices a wafer into hundreds of chips; Cerebras leaves it whole and makes one wafer-sized processor. The payoff is on-chip memory: 44 GB of SRAM on a single wafer.

Why that translates to speed becomes clear once you locate the bottleneck. Running a large language model on GPUs, the constraint is rarely compute — it's memory bandwidth. Generating each token means reading model weights out of memory, and that round trip dominates the time. In the Cerebras design, weights sit on the chip, and the round trip disappears.

OpenAI brought the model and the distribution. GPT-5.6 Sol is its top reasoning model, scoring 61 on Artificial Analysis's Intelligence Index, just under Claude Opus 5 (63) and Fable 5 (62). Putting that model on new hardware and selling it as a new API tier is the shape of this announcement.

Why speed matters now is that usage patterns changed. When a human asked a question and read the answer, 30 versus 60 tokens per second barely registered — reading is slower than that anyway. Agents broke the symmetry. An agent chains dozens to hundreds of model calls into a single task, and the latency of each call multiplies. Tokens per second becomes time-to-completion.

OpenAI researcher Jeffrey Wang's comment captures the shift: "It finishes for me before I even have the opportunity to context-switch. It makes me way more productive." Finishing before the context switch happens is what this speed regime actually buys.

What the numbers show

Here are the published figures.

Item Figure
Output speed Up to 750 tokens/second
Versus standard processing Up to 14x
Versus Claude Fable 5 11x
Versus Opus 4.8 on Fast mode 5x
HLE 2,500 questions — Ultrafast 11h 11m
HLE 2,500 questions — Claude Fable 5 78h 27m (7x gap)
GDP-Val end-to-end speedup 5.6x (no quality loss)
On-chip SRAM per wafer 44 GB
Delivery New OpenAI API service tier
Current status Limited preview, select customers
Announced August 13, 2026

The most contestable row is "no quality loss." Claiming a 14x speedup on the same model needs evidence, and what OpenAI offers is the GDP-Val result: 5.6x faster end to end with quality metrics held. If the speedup comes purely from swapping hardware, identical output is theoretically what you'd expect.

The HLE comparison deserves a careful read. 11 hours versus 78 is a dramatic contrast, but it blends two variables: hardware speed and how many tokens each model spends per question. Reasoning models think at length before answering, and that length differs by model. So the accurate reading isn't "Cerebras is 7x faster" — it's "this combination finished the same exam 7x sooner."

Access is worth flagging too. Ultrafast is delivered exclusively through the OpenAI API and is currently a limited preview for select customers, expanding as capacity grows. Nobody can just switch it on today. Wafer-scale chips don't ship in GPU volumes, so that constraint likely persists for a while.

The targeted use cases are named in the announcement: financial research, incident response, customer support, voice applications, commerce, and live experimentation. The common thread is obvious — in every one, something or someone is waiting. Batch work you can run overnight isn't on the list.

Who gains from this

OpenAI gains a reset axis of competition. The company is under pressure from two directions: above, Anthropic's Claude Opus 5 and Fable 5 lead on benchmarks; below, Grok 4.6 serves an equal score at one-fifth the price. Not first on intelligence, not first on price — so it created a third axis and took first place on that one.

Cerebras got the best possible reference. Wafer-scale architecture has long been treated as interesting but commercially unproven. The most-used frontier model running 14x faster on your hardware is the kind of evidence that ends that debate.

Companies building agent products are the practical beneficiaries. The number one complaint about agent UX is that they're slow. While a response is pending, users switch windows, and once they switch, the session dies. At 750 tokens per second, much of that dead zone disappears.

For NVIDIA it's a nuanced signal — not an immediate threat, since Cerebras production volumes aren't comparable. But this announcement is the strongest empirical case yet that inference workloads may suit a different architecture. Unlike training, the inference market has an open architecture race.

For users, there's no direct benefit yet. It's a limited preview, so most can't touch it, and pricing hasn't been disclosed. But once the tier opens up, products like real-time voice and live coding assistance will feel noticeably different.

We've seen these speed races before — the results split

Dedicated hardware dramatically accelerating a specific workload is a recurring story in semiconductors.

The archetypal success is the GPU itself. A chip built for graphics rendering turned out to suit parallel computation, and became the substrate of deep learning. The logic that worked was: move something off general-purpose processors onto a purpose-built structure and you get an order-of-magnitude improvement. Cerebras's argument has the same shape — with the difference that the GPU ecosystem is complete and Cerebras's isn't.

The second success is Google's TPU. Building custom silicon for its own workloads won on both cost and speed, and it's why Google has room to maneuver in model price competition today. The lesson there is that the real value of custom hardware is unit economics, not benchmark numbers.

The failure pattern is equally clear. Many AI chip startups posted impressive benchmarks and never found a market. Usually for one of two reasons: the software ecosystem never caught up, or model architectures shifted and the optimizations became irrelevant. Cerebras partnering directly with a frontier model company is a clever way around the first problem — instead of building an ecosystem, it borrowed one.

A third reference point is the shelf life of speed itself. A speed advantage is relative and narrows when competitors invest in the same direction. A price advantage comes from cost structure and tends to last longer. Whether this 14x gap holds for a quarter or several is unknown.

How the rivals counter

Anthropic answered on a different axis. A day later, on August 14, it made Claude Code's auto mode the default on paid plans — cutting perceived completion time by removing the approval waits rather than raising token throughput. And since a 78-hour HLE figure was just cited in a competitor's launch, a response on the speed axis wouldn't be surprising.

Google already owns TPUs, so it's free on the hardware axis. Its counter is a separate speed-oriented model line, like Gemini 3.5 Flash. And the billion monthly users it announced on August 11 is itself the economic justification for inference optimization.

SpaceXAI competes on price. Grok 4.6 serves the same benchmark score at one-fifth the cost, but its output speed is around 65 tokens per second — slower, not faster. Having closed the Cursor acquisition on August 14 and secured its own compute, how it moves on this axis is worth watching.

Other inference-hardware companies just watched the market open. A frontier lab formally adopting a non-GPU architecture raises credibility for the entire category. There's only one seat next to OpenAI, though.

NVIDIA will answer with inference-optimized product lines. Its dominance in training isn't wobbling, but inference is entering a phase of workload-specific architecture competition.

So what actually changes

If you build agents, your design assumptions may shift. Most agent UX today is designed around slowness — progress indicators, background execution, streaming intermediate results. In a 750-tokens-per-second regime, much of that scaffolding stops being necessary.

If you build real-time products, new possibilities open. Voice conversation, live coding assistance, incident response — domains where latency is product quality. Right now it's a limited preview, so whether you can act on it depends on getting access.

If you manage costs, there isn't enough information to judge yet. Pricing hasn't been published. Dedicated hardware tiers usually carry a premium, but finishing the same task much faster can also reduce total token consumption. You'll only know from an actual invoice.

If you follow AI infrastructure, the real significance is that the inference market's structure just opened. Training is effectively dominated by one architecture, but inference workloads vary enough that several architectures can coexist — and inference keeps taking a larger share of total compute demand.

For the industry, the headline is that there are now three axes: intelligence, price, and speed. No single company leads all three. Which means for the next several quarters, choosing a model becomes less about "the best model" and more about picking the axis your task actually needs.

🥄 Three Things You're Probably Wondering

— Same model, really the same quality? OpenAI reports no quality loss on GDP-Val alongside the 5.6x end-to-end speedup. If only the hardware changed, identical output is theoretically what you'd expect. The published validation is still narrow, though, so checking against your own workload is the reliable move.

— Can I use it now? Most people can't. It's delivered as a new OpenAI API tier and is currently a limited preview for select customers, expanding as capacity grows. Given wafer-scale production volumes, that constraint probably sticks around.

— Will every model get this fast? Too early to say. This speed comes from a specific hardware architecture with limited supply, and speed advantages narrow when competitors invest the same way. What does look permanent is that inference speed is now an established axis of competition.

References

Numbers are as of announcement and may change.