Tibo: Today's Ultra Fast at 750 token/s could become the norm in a year or two

Dynamic Beating AI News Flash: OpenAI now offers a paid high-speed tier called Ultra Fast, where you can pay to make the model respond even faster. Once activated, the GPT-5.6 Sol model can reach up to 750 Tokens per second, around 14 times faster than the regular mode. Codex lead Tibo Sottiaux estimates that in another 1 to 2 years, today's speed may be close to the default level.

However, the 14x speedup is only for the model's maximum output speed itself. When it comes to actual coding tasks, if the majority of time is spent generating code, the overall speedup is around 10x; if the Agent also needs to search the web, run tools, and wait for network responses, the overall task usually only speeds up by 3 to 4 times. As the model's output rate increases, the bottleneck shifts from the model to the network, CPU, and tool invocations.

Even within OpenAI, the Ultra Fast tier needs to be used sparingly. Major outages, critical prototypes, and similar tasks can take priority, but most of the computing power is reserved for external clients; otherwise, OpenAI's own staff could consume all the capacity.