OpenAI has previewed Ultrafast, a new API service mode designed to run GPT-5.6 Sol much faster without switching to a smaller model. The company says the Cerebras-powered mode can reach up to 14 times standard processing speed and generate up to 750 output tokens per second.

What OpenAI announced

The announcement appeared in OpenAI’s official feed on August 13 under the title “Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed.” OpenAI frames it as an API service tier for GPT-5.6 Sol, not a separate model. The core pitch is more useful work per second for tasks where latency changes how an AI system can be used.

OpenAI also published a YouTube video on August 14 with user testimony about the experience. In the transcript, participants describe incident investigation, internal-channel analysis, code refactoring and workflows where moving from one or two hours to 10 or 15 minutes changes the operating rhythm. The video is not a substitute for technical benchmarking, but it shows how OpenAI wants to position Ultrafast: infrastructure for near-real-time decisions and actions.

Cerebras' role

OpenAI says Ultrafast is powered by Cerebras. TechCrunch independently reported on August 13 that the preview relies on OpenAI’s partnership with the chipmaker and is currently available to a small group of customers, with broader access expected as capacity grows. That limitation matters: the announcement does not mean every API developer can immediately switch it on or that the advertised speed will hold across every workload.

The headline number — up to 750 tokens per second — points to use cases where generation speed becomes a product feature. In support, market analysis, incident response or software development, seconds can affect how many hypotheses are tested, how many documents are reviewed and how much coordination a team can do through an AI system.

Where it could matter

OpenAI and TechCrunch point to enterprise workflows such as incident response, customer support, financial-market analysis, e-commerce and software development. The common thread is not just doing the same work faster, but making powerful models usable in high-pressure interactive settings where slower responses were previously a constraint.

There is also a competitive infrastructure angle. AI labs are increasingly offering more capable, cheaper and faster model modes for different segments. Ultrafast pushes that contest toward serving infrastructure: if a frontier model can respond at speeds closer to specialized systems, companies may be able to use it in more operational processes without separating the “smart” model from the “fast” model.

What remains unproven

Confirmed facts include OpenAI’s official announcement, the August 14 official video, the claim of up to 14x speed, the up-to-750-output-tokens-per-second figure and Cerebras’ role. Still unproven publicly are performance across diverse real workloads, pricing, availability limits, access conditions and whether quality remains consistent across task types.

The right framing is therefore cautious: Ultrafast is a meaningful signal for enterprise AI — powerful models with much lower latency — but it is still a controlled preview, not a universally available capability or a guarantee of production results.

Sources: OpenAI, OpenAI’s official YouTube channel and TechCrunch.

Written by Nova Rivera — Product and automation perspective.