Codex Ultrafast: When AI Answers in Under 30 Seconds, Everything Changes

OpenAI has quietly shipped a mode called Ultrafast inside Codex, its AI coding agent, and their own engineers are already using it as their default when things break. The short version: responses that used to take a few minutes now land in under 30 seconds, and that small gap turns out to matter a lot.

Speed is not just a comfort feature

It's tempting to file "faster AI" under "nice to have." Don't. The engineers at OpenAI who use Codex daily are clear that the speed change is a behaviour change, not just a convenience upgrade.

One engineer, Adam Perry, put it plainly: before Ultrafast, he would ask Codex a question and then start investigating the same problem himself in parallel, because waiting felt like wasted time. Now he just waits. That is a fundamentally different way of working. Another engineer, Jeff Niu, says if an AI answer takes more than 30 seconds, he switches to Slack or asks a colleague instead. Ultrafast keeps most of his work under that threshold, which means Codex stays in the loop rather than getting bypassed.

For anyone building vibe coded projects with AI tools, this is the real lesson. The bottleneck was never only capability. It was the interruption. Every time you switch context while waiting for a response, you lose your thread. Cut the wait and you stop losing it.

What Ultrafast actually does

Ultrafast is a response mode in Codex, not a separate product. It is designed for situations where speed matters more than a long, deliberative answer, like incident response, quick lookups, or rapid back-and-forth debugging.

At OpenAI, engineers use it to run multiple lines of investigation at the same time rather than one at a time. AJ Trbojevic, who works on datacenter security, says pulling together structured and unstructured data during a security investigation used to take one to two hours. With Ultrafast, that same process takes 10 to 15 minutes, and sometimes happens close to real time. Sayan Sisodiya uses it to dig through Slack history alongside logs, looking for references to recent changes or people reporting similar symptoms elsewhere in the organisation.

For a designer who is learning to build, you are probably not running security incidents. But you are absolutely bouncing between a question, a long AI wait, and a forgotten train of thought. That is the same problem, just quieter.

How a designer could put this to work today

Here are a few concrete places where Ultrafast-style speed changes the workflow for someone building with Codex.

Debugging in real time. When something in your project breaks, you want to ask a question and stay with the problem. If you have to wait and switch tabs, you lose the mental picture of what went wrong. A faster response loop means you can follow one idea all the way through before jumping to another.

Keeping a running conversation. Adam Perry keeps a persistent Codex thread he treats like a chief of staff, asking followup questions on top of context that is already loaded. Because Codex does not have to re-fetch information each time, the conversation feels more like talking to a collaborator than submitting a ticket. You can build your own version of this: a thread you return to each day for status, next steps, and quick lookups.

Delegating browser tasks you kept putting off. Sayan Sisodiya mentions using Codex's computer use capability to fill out forms in a browser, something he had avoided before because the older, slower model would click, pause, think, and click again. With the latest models and faster response times, he is comfortable handing that work over. If you have repetitive web tasks sitting on your to-do list because they felt too fiddly to automate, this is worth revisiting.

Tackling refactors you deferred. Jeff Niu says technical debt is easier to pay down now because a refactor that used to feel expensive, in time and mental energy, now happens fast enough to be worth doing. For a designer managing their own codebase, that means the cleanup tasks you have been avoiding might finally be practical.

A grounded read on what this is and is not

Ultrafast is an internal tool at OpenAI, shared as an illustrative example of how the company uses its own technology. It is not a consumer product announcement, and there is no public pricing or release date in the source material. What OpenAI has shared is a set of genuine practitioner observations, not a polished case study, which makes it more useful as a signal.

The open question is how much of this speed advantage translates to the tools designers actually use today, whether that is the Codex API, ChatGPT, or a third-party app built on top. Response speed varies by model, plan, and how busy the servers are. The 30-second threshold Jeff Niu describes is a useful benchmark to hold in your head, though. If the AI tool you are using consistently takes longer than that for simple questions, it is worth testing alternatives, or checking whether a faster mode is available.

The core idea here is not about Codex specifically. It is about the relationship between speed and ambition. When the gap between question and answer shrinks, you ask more questions. You try more things. You stop avoiding the tasks that used to feel like a slog. That is a real shift, and it is already happening.