OpenAI previews Ultrafast: GPT-5.6 Sol at 750 tokens per second, powered by Cerebras

On 13 August 2026, OpenAI announced a limited API preview of Ultrafast, a new service tier that runs GPT-5.6 Sol at up to 750 output tokens per second. That is roughly 14 times faster than standard processing, which sits at around 53 tokens per second. The hardware behind it is Cerebras, whose wafer-scale chips make this kind of throughput possible without any reduction in model intelligence.

The headline number is striking, but the more important point is this: you are getting the full GPT-5.6 Sol, not a distilled or cut-down version. The speed comes from the hardware, not from simplifying the model.

Why is it this fast?

Standard GPU-based inference has a fundamental bottleneck: every time the model generates a token, it has to shuttle weights between off-chip memory (DRAM) and the processor. On large frontier models, that data movement is the rate limiter, not the arithmetic.

Cerebras takes a different approach. Its Wafer-Scale Engine 3 is a chip the size of an entire silicon wafer, at 46,225 mm², roughly 57 times the die area of an Nvidia H100. It holds 44 GB of on-chip SRAM, which is ten to twenty times faster than the DRAM used in conventional GPU memory. By keeping model weights on-chip, Cerebras eliminates the memory-bandwidth bottleneck almost entirely. The result is the kind of throughput that previously required trading model quality for speed.

OpenAI committed to $10 billion in Cerebras compute earlier in 2026. Ultrafast is the first public-facing product built on that capacity.

What the benchmarks show

Cerebras ran GPT-5.6 Sol on Ultrafast through Humanity’s Last Exam, a 2,500-question benchmark designed at PhD-level difficulty. Ultrafast completed all 2,500 questions in 11 hours and 11 minutes. Claude Fable 5 took 78 hours and 27 minutes on the same task. On GDP-Val, a benchmark for economically valuable knowledge work, Cerebras reports a 5.6 times end-to-end speedup over standard processing with no quality degradation.

For a direct model-to-model comparison: based on output speeds reported by Artificial Analysis, Ultrafast is 5 times faster than Claude Opus 4.8 in Fast mode, and 11 times faster than Claude Fable 5.

Who is already using it

OpenAI opened the preview to a small group of customers. The early adopters give a clear picture of what the tier is designed for.

Podium’s voice AI team noted that the speed “completely changes the call experience for the more complex work.” Jane Street said the throughput enables new working methods. Rogo, which focuses on financial research, described Ultrafast as making “complex financial research feel like a real-time interaction.” Basis framed it well: “Oftentimes the barrier to truly fast products is not just tokens per second, but also model intelligence, and Ultrafast combines both.”

OpenAI’s own incident response team uses it internally to read logs, analyse traces, synthesise conversations, and help prepare or validate fixes. The company also notes it tightens the research iteration loop: rather than launching experiments overnight and reviewing results in the morning, teams can run multiple iterations during the working day.

What this means for you

The use cases OpenAI has identified are a useful guide to where Ultrafast actually earns its keep: real-time voice agents, customer support, incident response, financial analysis, and developer agents operating in production.

If your application requires a human-feeling response time, standard inference at 53 tokens per second means a 400-word reply takes roughly 15 seconds. At 750 tokens per second, the same reply arrives in just over one second. That difference matters enormously in voice, live support, and any workflow where a person is waiting on screen.

It matters less for batch processing, asynchronous summarisation, or tasks where the user is not watching a timer. If that describes most of your workload, the existing Fast tier (which runs GPT-5.6 Sol at up to 2.5 times standard speed) may be the better fit.

For teams building agents that need to reason through complex problems in real time, the combination of frontier-model intelligence and this level of throughput removes a constraint that previously forced a choice between quality and speed.

Availability and how to get access

The Ultrafast tier is available in the OpenAI API as a limited preview from 13 August 2026, with access expanding as Cerebras capacity grows. Businesses can join the waitlist by sharing their workload, latency requirements, and expected usage.

Pricing has not been publicly announced for the preview period.

The bigger picture

This launch sits within a broader layered compute strategy at OpenAI: Nvidia H100 and H200 GPUs for training and general serving, Cerebras wafer-scale chips for latency-critical inference, and a future cost-optimised tier (codenamed Jalapeño) for production serving at scale. Ultrafast is not a replacement for that stack; it is a specialised layer within it.

For Cerebras, which went public in one of 2026’s larger listings but has faced scepticism about whether wafer-scale economics translate into durable margins, serving OpenAI’s flagship model in production is a meaningful reference point.

Sachin Katti, VP Compute Strategy and GPT-Infra at OpenAI, put it plainly: “By combining GPT-5.6 Sol with Cerebras’ inference technology, we’re exploring what becomes possible when customers can get the intelligence of our most capable models with significantly lower latency.”

That framing is the right one. The question Ultrafast is trying to answer is not “how fast can a model go?” but “what products become possible when the model is fast enough to feel instant?” The early customer list suggests some of those answers are already emerging.