All the models that are fit to print.
OpenAI · Anthropic · xAI

The Daily Model

At OpenAI, Codex Got Fast Enough to Wait For


OpenAI's own engineers say Codex's Ultrafast mode cut incident investigations from hours to minutes and kept their attention in one place; the figures are self-reported, and among…


Jeff Niu, a Member of the Technical Staff at OpenAI, has a rule for working with AI. If an answer takes more than 30 seconds, he checks Slack or chats with a coworker. He used to ask Codex a question, switch to something else while it thought, and come back having lost his train of thought. Most of his work, he says, is now under that 30-second threshold.

That account comes from a case study OpenAI published on October 26, 2026, about Ultrafast, a high-speed mode in Codex (openai.com). It is part of the company's "OpenAI on OpenAI" series of internal examples, and OpenAI says Ultrafast has become the default for incident response among its engineers.

The figures are specific. AJ Trbojevic, a Member of the Technical Staff and a Datacenter Security Architect, says security investigations used to take an hour or two to pull data together, normalize it and provide context to make a decision. "Now that happens in 10 to 15 minutes," he says. "And at times it's near real time." Aaron Friel, who usually runs point on incidents, says Ultrafast lets him "essentially search all of those haystacks at the same time" (openai.com).

What changed most, in their telling, is attention. Adam Perry, who works on the core agent team, says he used to ask Codex to look into something and then look into it himself at the same time, because he knew it wouldn't be fast enough to wait on. "Now with Ultrafast, Codex is fast enough that I can just stay focused instead of splitting my attention," he says.

Ultrafast is not only an internal tool. OpenAI's release notes, checked on October 27, 2026, describe it as a faster service tier for GPT-6 Astra, chosen in the model picker in ChatGPT Work and Codex. Among ChatGPT's Pro plans, only Pro 500, at $500 a month, includes it; buying credits on Pro 100 or Pro 200 does not unlock it (help.openai.com).

Every figure in the case study comes from OpenAI staff, in OpenAI's own publication, and the post offers no independent measurement. It describes how a few named engineers work, not how fast Ultrafast will be for anyone else.

Still, the observation underneath it is worth keeping. For Niu, the line sits at 30 seconds. Under it, the tool stays part of his thinking. Over it, the question becomes something to check on later, after the thought has gone.

Sources

Read elsewhere

Published in the edition of Tuesday, October 27, 2026.