OpenAI is previewing Ultrafast, a new service tier for its OpenAI API that the company says can run GPT-5.6 Sol at up to 14 times the speed of standard processing. Powered by chipmaker Cerebras, the mode is designed to deliver as many as 750 output tokens per second and bring a high-capability model into workflows where response time directly affects the outcome.

The launch is starting as a limited preview for a select group of customers. OpenAI says it will expand access as capacity grows, making availability—not just model performance—the immediate constraint for businesses evaluating the service.

What OpenAI is launching

Ultrafast is not presented as a separate model. It is a new speed tier for GPT-5.6 Sol, which OpenAI describes as its most intelligent model. The company says the tier achieves up to a 14-fold improvement over Standard processing, with Cerebras providing the inference infrastructure behind the preview.

OpenAI’s announcement focuses on latency rather than a new reasoning capability or expanded knowledge cutoff. Its argument is that companies have traditionally had to choose between more capable models and faster, smaller, or specialized systems. Ultrafast is intended to reduce that tradeoff by allowing GPT-5.6 Sol to respond quickly enough for interactive applications.

The product is launching first through the OpenAI API, rather than as a broadly available ChatGPT feature. That distinction matters for builders: the initial opportunity is to embed faster model responses into software, operational tools, and customer-facing systems, while access and capacity remain controlled by OpenAI.

Where the speed could matter

OpenAI is testing Ultrafast across business workflows including incident response, coding, commerce, financial research, customer support, and other interactive applications. In an incident response setting, the company says engineers can use the model to review logs, traces, recent code changes, and internal conversations while an outage is still developing. The model can help identify likely causes, suggest checks, and prepare or validate a fix, but engineers remain responsible for judgment and deployment.

The same timing advantage could change the design of customer support and voice products. OpenAI says faster responses could help resolve complex issues during a live conversation, even when the answer requires consulting multiple systems. In commerce, it highlights product questions, inventory checks, recommendations, and checkout assistance as tasks that may lose value if a customer has to wait.

OpenAI also points to research workflows. Teams that currently launch experiments overnight and review results the next day could, in principle, run more iterations during working hours. That does not remove the need for data preparation, experiment design, or evaluation, but it could shorten the loop between a question, a result, and a revised approach.

Evidence and open questions

The central performance figures are vendor-reported. OpenAI says Ultrafast can reach 750 output tokens per second and operate up to 14 times faster than Standard processing, but the announcement does not provide an independent benchmark, a latency distribution, a representative prompt set, or details on how throughput changes with input length and concurrent demand.

The figures also describe output speed, which is only one part of the user experience. Time to first token, tool-call latency, retrieval delays, queueing, and the time required for a complete multi-step task will determine whether an application actually feels 14 times faster. OpenAI has not disclosed pricing, capacity commitments, or the operating limits that will apply during the preview.

Adoption evidence is similarly early. OpenAI says it is working with an initial group of companies and that teams inside the company are testing the service. Those examples demonstrate intended use cases, not independent proof of production reliability, cost efficiency, or improved business results. TechCrunch AI placed the launch in a wider race to accelerate model responses, noting that Anthropic has also offered a fast mode for Claude, although the two services have not been compared through a shared evaluation in the available evidence.

Implications for builders and enterprises

For product teams, the most important question is not whether a model can generate text quickly in isolation. It is whether faster inference changes a workflow enough to justify its cost and integration complexity. A support agent that waits on retrieval, policy checks, and account permissions may still be slow even if the model itself produces 750 tokens per second. Builders will need to measure the full request path and determine whether speed improves resolution rates, conversation quality, or user retention.

The preview could be especially relevant to systems that currently use smaller models to meet latency budgets. Replacing those models with GPT-5.6 Sol may simplify architecture if quality improves without unacceptable cost or throughput constraints. But enterprises will still need safeguards for financial analysis, incident response, and customer communications, where a rapid incorrect answer can increase risk rather than reduce it.

Cerebras is also becoming a material part of the product story. OpenAI’s partnership suggests that model capability and inference hardware are increasingly being packaged together as differentiated service tiers. For buyers, that makes infrastructure availability, regional deployment, data handling, uptime, and predictable capacity important evaluation criteria alongside benchmark quality.

What to watch next

The next meaningful signals will be broader access, published pricing, and independent measurements of time to first token and end-to-end task latency. Buyers should also look for evidence about throughput under concurrent enterprise workloads, rate limits, tool-use performance, and whether Ultrafast maintains GPT-5.6 Sol’s quality on long-context and multi-step tasks.

OpenAI’s expansion timeline will reveal whether Cerebras-backed capacity can support more than a small preview group. Case studies with measurable outcomes—such as shorter incident resolution, higher support containment, or faster research iteration—would be more useful than raw token throughput alone.

Creati.ai perspective

Ultrafast is significant because it treats inference speed as a product capability rather than a backend optimization. If OpenAI can make a frontier model responsive enough for live operational and customer workflows, developers may redesign applications around continuous interaction instead of queued requests and overnight jobs.

The launch remains an early capacity and validation story. Until pricing, availability, independent benchmarks, and production results emerge, the 14-times figure should be treated as an OpenAI claim about peak service performance—not as proof that every enterprise workflow will become 14 times faster.