OpenAI has introduced a limited preview called Ultrafast for GPT-5.6 Sol, claiming that it can process requests 14 times faster than the model’s standard mode and produce as many as 750 output tokens per second. The capability is powered by the company’s partnership with Cerebras and is initially available only to a small group of customers.
Speed has become a product characteristic in its own right. A capable model may still feel unsuitable for an interactive workflow if users wait too long for an answer, especially when the model sits inside a repeated operational process. Ultrafast is OpenAI’s attempt to reduce that delay without presenting it as a different underlying flagship model.
The announcement is a preview, not broad availability. OpenAI says access will expand as capacity grows, but it has not provided a general release date in the source report. The headline performance figures are also company claims; independent results across varied prompts and workloads are not included.
Faster output changes which workflows feel interactive
OpenAI identifies incident response, customer support, financial analysis and ecommerce as possible uses. Each example involves decisions or conversations where waiting can interrupt attention or slow a sequence of actions. Higher token throughput could let a user review a long response sooner or allow software to complete several model-dependent steps in less time.
The practical effect will depend on more than the maximum number of generated tokens per second. A request must first reach the system, be prepared and begin producing an answer. The TechCrunch report does not provide measurements for that complete journey, nor does it compare performance across different request sizes. The claimed 14-fold improvement should therefore be read as OpenAI’s description of the preview rather than a universal expectation.
Even so, a faster mode can change product design. Developers may be able to use a strong model in experiences that previously required a trade-off toward a quicker option. Users may also be more willing to iterate when each attempt returns promptly. Those possibilities remain projections until customers test the service under realistic load.
Cerebras supplies the infrastructure behind the preview
Ultrafast is powered through OpenAI’s partnership with Cerebras. That attribution makes the launch an infrastructure story as much as a model update: the software capability is being differentiated through the hardware and serving system that deliver it.
The preview also indicates that frontier AI products may offer multiple performance profiles around the same model. OpenAI is not alone in pursuing faster access; TechCrunch notes that Anthropic offers a fast mode, although the source says it does not reach the same speed. No detailed like-for-like benchmark is provided, so the comparison cannot establish a definitive ranking.
Capacity is the most visible constraint. Limiting Ultrafast to a small customer group allows OpenAI to observe demand and performance before wider deployment. It also means that the announced speed is not yet something most customers can select. Expansion will depend on the infrastructure available through the partnership and on how OpenAI allocates that capacity.
A preview must prove more than its peak speed
For customers, the relevant question is whether lower latency produces better operational results. A support team might resolve conversations sooner, while an incident-response team might analyse material without long pauses. But speed does not guarantee answer quality, factual reliability or safe action. Those properties still require evaluation for each use case.
The source report does not disclose pricing, service guarantees, regional availability or how Ultrafast behaves during sustained demand. Those missing details will shape whether the mode is appropriate for occasional premium tasks or routine high-volume deployment. Organisations also need to determine whether a faster response is valuable enough to justify any eventual cost difference.
Ultrafast demonstrates the direction of travel: frontier-model competition is extending beyond intelligence scores into the experience of receiving an answer. If OpenAI can reproduce its claimed throughput at scale, sophisticated models could feel less like tools that users consult and more like responsive components embedded in live work. For now, the small preview provides a performance claim and a test bed—not yet proof of broad, dependable real-time AI.
