OpenAI has launched a new Ultrafast mode for its flagship GPT-5.6 Sol, promising to increase processing speed fourteenfold. This improvement goes beyond a simple performance tweak; it aims to eliminate the traditional trade-off between speed and capacity, delivering near-instantaneous responses without compromising the intelligence of the company's most advanced model.
The new technology, currently available in preview to a select group of customers, leverages the infrastructure of Cerebras, a chip manufacturer specializing in AI inference. According to data provided by OpenAI, the system can generate up to 750 tokens per second , a figure four times greater than most current commercial models and positioning this technology at a new level within the industry.
A leap in speed without sacrificing the model
Until now, achieving a fast response usually meant resorting to a smaller or more specialized model, a compromise OpenAI has sought to avoid with Ultrafast. Instead of creating a stripped-down version, the company has opted to accelerate the same GPT-5.6 Sol , leaving its reasoning and text generation capabilities intact. This means that developers and companies using the API will be able to enjoy minimal latency without sacrificing the quality of the results.
The speed of 750 tokens per second means that a paragraph of response is produced in less than a second, making interaction virtually instantaneous. This feature is especially valuable in applications where every millisecond counts, such as algorithmic trading, customer support with high volumes of inquiries, or security incident response. OpenAI emphasizes that speed no longer requires sacrificing intelligence , a message that has resonated with early preview users.
Cerebras' infrastructure, key to performance
The secret behind Ultrafast lies in its collaboration with Cerebras, whose Wafer-Scale Engine processor integrates 44 GB of SRAM and keeps the model weights close to the compute units. This architecture drastically reduces data movement between external memory and the processor, one of the main bottlenecks in inferring large models. According to the company, the weights remain on the chip while the tokens traverse the layers , allowing it to handle hundreds of tokens per second without degradation.
The integration with Cerebras not only benefits OpenAI but also marks a milestone for the chipmaker, which sees its infrastructure supporting a frontier model within a commercial service. While capacity remains limited, both companies have confirmed they will expand access as resources become available. Currently, Ultrafast is offered only via the API and is not available on ChatGPT, a distinction OpenAI has made clear in its press release.
Use cases designed for critical environments
OpenAI has identified several scenarios where low latency makes a significant difference. These include incident response , where an agent can analyze logs, review traces, and propose solutions while the outage is still active; financial research , which requires processing market signals in real time; and customer service , where conversations flow seamlessly without perceptible delays. E-commerce is also mentioned, with personalized recommendations and live inventory management.
Companies like Jane Street, Podium, Basis, and Rogo are already testing Ultrafast in applications related to programming, commerce, and support. Initial results indicate that reducing the time spent on each task can transform processes that previously ran in nightly batches into continuous iterations throughout the day . However, OpenAI cautions that final performance depends on the complexity of each workflow, including external tools and database calls, so the figure of 750 tokens per second is only a reference point for the maximum generation rate.
Limited availability and next steps
Ultrafast is currently in a restricted preview phase, and OpenAI has not announced a date for its general availability. The company has confirmed that access will expand as capacity increases , but for now, only a select group of customers can activate it. The public price of this new service level has also not been revealed, a crucial factor for businesses to assess whether the reduced latency justifies the additional cost.
Compared to the competition, OpenAI claims that Ultrafast outperforms Anthropic's Claude Fast mode, which doesn't achieve the same speeds. This difference could become a deciding factor for organizations that rely on near-instantaneous responses. Meanwhile, developers are already envisioning applications that were previously impractical due to wait times, from real-time code generation to interactive text-based games.
In short, the arrival of Ultrafast marks a turning point in the race for the fastest AI on the market. With the collaboration of Cerebras and an approach that doesn't compromise on the model's power, OpenAI has clearly established a benchmark in the realm of low latency. Although pricing and general availability are still pending, the preview is already demonstrating that speed can go hand in hand with intelligence , and that the seconds saved on each task have real value in enterprise environments.