OpenAI Ultrafast Mode: GPT-5.6 Sol at Up to 14x Speed, Powered by Cerebras

On August 13, 2026, Cerebras announced that it is powering Ultrafast mode, a new service tier in the OpenAI API¹. It targets GPT-5.6 Sol and runs it at up to 750 output tokens per second and up to 14x faster than Standard processing¹. Availability is initially limited to a preview for OpenAI customers¹.

The notable part is that this is not a model swap. Cerebras states that GPT-5.6 Sol on Ultrafast runs with the same intelligence as GPT-5.6 Sol Standard¹. The familiar tradeoff — drop to a smaller or specialized model when you need speed — may take a different shape.

What Is Being Offered

Ultrafast mode is positioned as a service tier within the OpenAI API¹. During preview, the customer set is deliberately small. Sachin Katti, VP Compute Strategy & GPT-Infra at OpenAI, is quoted in the Cerebras announcement saying the companies are starting with a small group of customers to learn where that speed creates meaningful value, and will use those learnings to inform how the service expands over time¹. Cerebras is collecting sign-ups for notification as capacity expands¹.

Neither the Cerebras announcement nor TechCrunch’s report mentions pricing. The cost difference relative to Standard processing has not been disclosed.

GPT-5.6 became generally available in July 2026 across three tiers — Sol, Terra, and Luna, with Sol at the top. On July 30, Luna received an 80% price cut and the API’s priority processing option was reworked into Fast mode, where Sol was offered at up to 2.5x Standard speed for twice the price. The new up-to-14x figure describes a separate tier along the same axis.

Where the Speed Comes From, and How to Read the Comparisons

The stated source of the speed is the Cerebras Wafer-Scale Engine architecture. It keeps model weights on-chip — 44 GB of SRAM on each wafer-sized chip — rather than shuttling them between on-chip memory and off-chip storage as GPU-based inference must, which the company says eliminates the memory-bandwidth bottleneck that constrains frontier-model inference speed on conventional hardware¹.

Comparisons against other vendors’ models are also published, and their provenance is worth noting. Cerebras says Ultrafast is 5x faster than Claude Opus 4.8 in Fast mode and 11x faster than Claude Fable 5 — but based on output speeds for Anthropic models reported by Artificial Analysis¹. These are not measurements Cerebras took itself under matched conditions.

Benchmark figures are included as well. On Humanity’s Last Exam, a 2,500-question benchmark spanning graduate-level chemistry, economics, and literature, GPT-5.6 Sol Ultrafast answered the full set in just over 11 hours¹. The Cerebras blog gives a finer figure of 11 hours and 11 minutes, against 78 hours and 27 minutes for Claude Fable 5². The framing is that it reached comparable accuracy nearly 7x faster¹. On GDP-Val, a benchmark of knowledge-work tasks such as legal briefs, financial models, and engineering reports, Ultrafast delivered a 5.6x end-to-end speedup with no loss in quality¹.

All of these are measurements supplied by Cerebras, not independently verified by a third party. Whether the same results hold on a given production workload is a separate question.

An Alternative to “Fast but Smaller”

Cerebras ties the announcement to an argument that speed, not capability, is now the binding constraint. The company’s framing is that organizations have until now had to choose between the capabilities of larger models and the faster response times of smaller ones¹. CEO Andrew Feldman says the release is proof that speed and intelligence are no longer mutually exclusive¹. TechCrunch likewise reports that getting real-time speed has typically meant choosing a smaller or more specialized model³.

The use cases cited include production incident response, cybersecurity threat detection, legal document analysis, and financial modeling², with TechCrunch adding customer service and support, financial market analysis, and e-commerce³. In each, latency is a constraint on the work itself.

The practical impact lands wherever latency is the bottleneck. Conversational UIs, agent loops that make dozens of tool calls, and code generation where someone is waiting on a review all feel per-token generation speed directly in both responsiveness and iteration count. For batch jobs and overnight processing that can afford to wait, there is little to evaluate while pricing remains unpublished. For how token generation relates to billing, see What Is a Token.

Pursuing inference speed through dedicated hardware is visible on the lab side too. Anthropic posted openings for an in-house chip design team in August, aiming at co-design of models and hardware. If frontier model offerings increasingly differentiate not just on model quality but on which hardware returns an answer in how many seconds, then choosing an API means comparing speed tiers as well.