OpenAI Speeds Up GPT-5.6 by Up to 14x for Tasks Where Seconds Matter

Language-model performance has often been judged mainly by answer quality and price. On August 13, OpenAI added a third parameter: time. The new Ultrafast service is intended to run GPT-5.6 Sol up to fourteen times faster than standard mode and generate as many as 750 output tokens per second. OpenAI is initially launching it in limited access through its API, with the compute layer provided by Cerebras. Technologically, this is a significant shift because developers seeking very low latency have often chosen a smaller model. Ultrafast is intended to make the same powerful model usable even where a system must react almost immediately. OpenAI cites log analysis during service outages, evaluation of suspicious financial transactions, voice support, and e-commerce, where several seconds of waiting can affect whether a purchase is completed. The effect on product design is particularly important. In ordinary office work, generative AI can take tens of seconds or minutes because the user is waiting for a document to be completed. During a customer call or operational incident, however, the system must respond alongside a human. If the model searches several systems, checks an order, and proposes a solution during the conversation, every delay directly affects the customer interaction. OpenAI describes a similar effect in research. A batch of experiments that a team launches in the evening and reviews the next morning can become closer to interactive work with a faster model: the researcher evaluates a result, changes a hypothesis, and runs several additional rounds within the same working day. In operations, the interval between detecting a signal, forming a hypothesis, and testing an intervention can likewise be shortened. The company has not yet published general pricing or a date for full availability. That is critical for any economic assessment. Fourteen times the speed does not automatically mean fourteen times the productivity. In many processes, the bottleneck is a person, a database, or an external system. The greatest benefit is therefore likely to arise where waiting for the model genuinely blocks the next step. Podium, for example, is testing Ultrafast in voice services. Product leader Courtland Lykins said: “The speed completely changes the call experience for the more complex work.” Rogo similarly says that complex financial research begins to feel like real-time interactive work when latency falls. For companies, this creates another parameter for model selection. Alongside price per token and answer quality, they will need to calculate the economic value of a second of waiting. In automated document processing, that value may be almost zero. On a customer line, in a commercial transaction, or during a security incident, it may be substantial. Key terms Latency: The time between sending a request and receiving the system's response. Cerebras: A manufacturer of specialized computing systems used, among other things, to run large AI models. API: A programming interface that allows other applications to communicate automatically with a service or model.

OpenAI Speeds Up GPT-5.6 by Up to 14x for Tasks Where Seconds Matter

The full article is available to users registered at Hard Skills