Skip to content
Main Site News Console

How Ultrafast Cuts OpenAI Troubleshooting Time from Hours to Minutes

· OpenAI Translated
OpenAILLM

October 26, 2026

OpenAI on OpenAI

This is part of our series on how OpenAI uses its own technology and APIs internally. These tools are actively deployed across OpenAI, and we’re sharing them to highlight how frontier AI powers diverse use cases across our teams.

Ultrafast is a high-speed mode in Codex that has become the default choice for OpenAI engineers during incident response. When systems fail and issues escalate, engineers can’t afford to wait minutes for answers. With Ultrafast, they can investigate multiple potential root causes in parallel and get answers without losing focus.

“If you have two equally smart people, and one takes an hour to respond while the other takes a minute, the second person is just strictly better. You get so much more done working with someone who responds faster.”

—Adam Perry, Member of Technical Staff at OpenAI

Finding the needle in the haystack

Aaron Friel, a Member of Technical Staff at OpenAI, often leads incident response efforts. He sets up an incident Slack channel and brings Codex into the loop, asking it to monitor, troubleshoot, and answer questions as the situation unfolds. “In any incident, there are usually many steps where things could go wrong, and the main job for responders in the channel is finding the right needle in the right haystack,” he says. “With Ultrafast, I can essentially search all of those haystacks at once, dramatically speeding up incident response.”

AJ Trbojevic, Member of Technical Staff and Datacenter Security Architect, notes that investigating security issues used to take an hour or two just to aggregate structured and unstructured data, normalize it, and provide the context needed to make decisions. “Now it takes 10 to 15 minutes,” he says. “Sometimes it’s almost real-time.”

Adam Perry and Sayan Sisodiya both work on the Core Agents team, writing the code that allows Codex to take actions, run commands, and use tools. Ultrafast has reshaped how they work during security incidents. “Previously, I would ask Codex to investigate something and then go look into it myself at the same time because I knew waiting around wasn’t an option,” says Adam. “Now with Ultrafast, Codex is fast enough that I can stay focused instead of splitting my attention.”

Sayan appreciates Ultrafast’s ability to pull answers instantly across multiple data sources: “During an incident, besides checking logs, I use Codex to search Slack to see if there’s any discussion around suspicious changes, or if anyone elsewhere in Slack is talking about similar anomalies,” he says.

Responses in under 30 seconds

Jeff Niu, another Member of Technical Staff, notes that if getting an answer from an AI takes longer than 30 seconds, he tends to wander off to check Slack or chat with coworkers. But with Ultrafast, most of his queries now resolve well within that threshold. Previously, he would ask Codex a question, switch to another task while it processed, and lose his train of thought by the time he returned. Now, the interaction happens in real time, helping him stay locked in on the task at hand.

Ultrafast has also made Jeff much more willing to tackle large-scale refactors, as rewriting existing code takes far less time. “Paying down tech debt is easier than ever,” he says. “Code is cheap, and it generates so fast now. Refactors that used to be painfully expensive can now be done almost instantly.”

Working together as a copilot

Adam keeps an ongoing Codex thread that he refers to as his “Chief of Staff.” Every morning, it briefs him on his schedule for the day and flags any unread messages he might have missed. “It already looked at all the data when preparing the summary, so when I ask follow-up questions, it doesn’t need to spend time refetching anything,” he says. “I can just keep asking questions in the same thread, like I would with an actual chief of staff.”

Sayan uses Codex to handle tedious, repetitive browser-based work like filling out forms—tasks he previously avoided delegating to AI because it was too slow. “It used to open the browser, pause to think for five seconds, click once, and pause again,” he says. With the latest models improving at Computer Use, he feels much more confident handing these tasks off to Codex. “Now I’m completely comfortable letting it drive the browser. It’s fast enough that delegating to it doesn’t waste any time.”

Whether for incident response, daily planning, or code refactoring, Ultrafast allows OpenAI’s technical team to work alongside Codex at the speed of thought. They can immediately ask follow-up questions, explore alternative approaches, and stay in flow without breaking their momentum. Shorter wait times have also changed the types of tasks engineers are willing to take on—from long-deferred code refactors to everyday chores that are finally worth delegating.