On August 13, 2026, OpenAI previewed Ultrafast mode for GPT-5.6 Sol as a new speed tier. It launches first through the OpenAI API and is powered by Cerebras inference hardware. OpenAI says Ultrafast can generate up to 750 output tokens per second, or as much as 14 times the speed of Standard processing. Those are vendor-published upper bounds; real performance will depend on the prompt, output length, tool calls, and load.
The point is not simply to choose a smaller, faster model with less capability. OpenAI is trying to put frontier intelligence into workflows that need an immediate response. Its examples include incident response, financial research and security analysis, real-time customer support, commerce conversations, and research experiments that can be iterated repeatedly during the workday. For an agent, lower latency can shorten the loop between observing a signal, calling a tool, checking the result, and deciding what to do next.
OpenAI highlights incident response as one use case. When a production system fails, an agent could quickly read logs, traces, recent code changes, and engineer reports to identify likely causes and prepare a fix. That remains assistance, not autonomous deployment. OpenAI says engineers stay responsible for judgment and deployment. The value of speed is that people can see enough evidence earlier, not that approval disappears.
For research, OpenAI says its internal teams use Ultrafast to search knowledge, query data, and organize results across connected tools. An experiment loop that once ran overnight could, in principle, support more iterations during the day. This is an internal workflow and an early observation described by OpenAI, not an independent study. Teams adopting it still need to measure quality, cost, and failure rates themselves.
Ultrafast is currently a limited preview for an initial group of customers, with access expected to expand as capacity grows. That rollout matters because faster inference does not automatically mean lower total cost, nor does it remove bottlenecks in external APIs, data permissions, or human approval. Looking only at tokens per second can hide the end-to-end latency of the whole agent workflow.
A useful evaluation should record time to first token, time to completion, tool-call waiting time, cost per task, answer accuracy, human edit rate, and rerun frequency. Customer support, transaction monitoring, and incident response also need peak-load tests, stability tests while data is changing, and explicit permission-boundary checks. Speed becomes useful productivity only when the result remains reliable, traceable, and approvable.
The broader signal is that AI platforms are starting to compete on useful work per second alongside model scores. As models handle longer tasks, latency affects how many attempts an agent can make, how long a user will wait, and whether an interactive product can preserve context. OpenAI's preview is early, but it points to the next AI workflow contest being fought across intelligence, speed, economics, and governance at once.
