OpenAI Previews Ultrafast Mode for GPT-5.6 Sol, Hitting 750 Tokens Per Second via Cerebras
OpenAI released a preview of Ultrafast, a new API service tier for GPT-5.6 Sol that runs the model at up to 14 times the speed of standard processing, delivering up to 750 output tokens per second. The preview is powered by OpenAI's partnership with chipmaker Cerebras and is currently available to a small group of customers, with access set to expand as capacity grows. OpenAI
The throughput figure is the headline number that matters for builders. At 750 tokens per second, Ultrafast operates in a regime where token generation latency approaches the point of being effectively imperceptible for interactive applications. For context, GPT-5.6 Sol's existing Fast mode delivers up to 2.5x faster speeds than Standard processing at twice the price, with no change in intelligence, according to OpenAI's earlier pricing and performance disclosures. Ultrafast sits in a different category entirely, jumping from 2.5x to 14x over the Standard baseline.
OpenAI is positioning Ultrafast for deployment across corporate workflows including incident response, customer service and support, financial market analysis, and e-commerce, according to TechCrunch. These are use cases where inference latency is not merely a UX concern but an operational constraint — an incident-response agent that can generate remediation steps in under a second changes the economics of automated operations, and a financial-analysis pipeline that produces output at 750 tokens per second can surface insights in near real time.
The model itself, GPT-5.6 Sol, is described by OpenAI as a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with the company's most advanced safety measures. It achieves state-of-the-art results across coding and knowledge work, per OpenAI's model announcement. An August 5 update to GPT-5.6 Sol in ChatGPT delivered more focused answers and adapted its level of detail to the question being asked, according to OpenAI.
The competitive landscape adds a useful frame. Anthropic's Claude offers a "fast mode," but it does not deliver the speed that Ultrafast offers, TechCrunch reports. The 14x multiplier, if it holds in production workloads, puts OpenAI in a distinctly different performance tier from what Anthropic currently ships.
The Cerebras partnership is the enabling factor. Cerebras builds wafer-scale inference hardware, and the architecture is well suited to the kind of sustained high-throughput token generation that 750 tokens per second demands. The fact that OpenAI is routing a premium model tier through a third-party chipmaker rather than running it on its own infrastructure is worth noting — it suggests that the computational profile of Ultrafast is best served by silicon that was not designed with GPU-style generality in mind.
OpenAI's August 13 announcement was accompanied by a companion publication, "The builder's guide to GPT-5.6," under its Applied AI category, suggesting the company is investing in developer enablement alongside the raw performance claim. The same day brought news of Dali Rajic's appointment as Chief Revenue Officer. Earlier in the week, OpenAI confirmed it is testing ads in ChatGPT (August 11), made Daybreak models available on AWS (August 11), and published two security posts on August 10 expanding its Daybreak cyber defense initiatives and putting frontier cyber models in more trusted hands. An August 12 company post titled "How enterprises put AI to work" rounded out a week of enterprise-facing messaging.
Looking at what this means in practice, the throughput jump from Fast to Ultrafast is not incremental. It crosses a threshold where the model's output speed stops being the bottleneck in most human-facing and many machine-facing pipelines. At that point, the constraints shift upstream — to retrieval, to orchestration, to the quality of the prompts and the context window management — and downstream, to how fast consuming systems can act on generated output. For engineering teams building on GPT-5.6 Sol, the relevant question is no longer whether the model is fast enough for a given use case but whether the surrounding architecture can keep up.
The limited preview scope is the practical caveat. OpenAI has not disclosed pricing for Ultrafast, nor a timeline for general availability, nor the specific throughput characteristics under sustained load versus burst. What is known is that the tier exists, it runs at 14x standard speed on Cerebras hardware, and it is being tested with a small set of customers. Builders evaluating whether to architect around it should treat those gaps as open variables.
