Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarOpenAI Blog→ original

OpenAI's Ultrafast Mode in Codex Cut Incident Investigations from 1–2 Hours to 10–15 Minutes

When something breaks in production, engineers do not have time to wait several minutes for an AI assistant's answer. That is why OpenAI's engineers made Ultrafast, a fast mode for the Codex AI agent that searches logs and chats for causes, the standard for incident reviews. A security investigation that once took one to two hours now fits into 10–15 minutes. For on-call teams, the takeaway is simple: the speed of the agent's responses determines whether people will use it instead of searching by hand.

AI-processed from OpenAI Blog; edited by Hamidun News
OpenAI's Ultrafast Mode in Codex Cut Incident Investigations from 1–2 Hours to 10–15 Minutes
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
Key metric #1Jeff Ngu's patience threshold is 30 seconds, after which he goes to Slack
Key metric #2Adam Perry used to wait for Codex while searching on his own in parallel; now he stays focused
Key metric #3Aaron Fril launches a dedicated Slack channel and connects Codex to it

Picture this: something breaks in production, users are already complaining, and the AI assistant takes several minutes to think through every question. While it answers, the engineer goes digging through the logs himself, and the agent's value drops to zero. OpenAI described how it solved this problem internally. Ultrafast, a fast mode for Codex, its AI agent for programming and working with tools, has become the default mode for engineers doing incident reviews. The headline figure: a security investigation that used to take one to two hours now takes 10–15 minutes, and sometimes happens almost in real time.

An important caveat: this is the company's account of its own tool. There are no independent measurements, and the material does not provide full figures on price or token consumption. Still, the scenario clearly shows how working with an agent changes when it responds in seconds.

What It Looks Like During an Outage

Aaron Fril, an OpenAI engineer, usually leads incidents. He creates a dedicated Slack channel and connects Codex: the agent is meant to monitor the situation, look for the cause, and answer questions as they come up. According to Fril, any outage has many possible links where something could have broken, and the main job of the people in the channel is to find the right needle in the right haystack. Ultrafast lets the agent search those haystacks simultaneously rather than one after another.

The most concrete numbers came from AJ Trboevich, a data center security architect. Previously, a security investigation took one to two hours: you had to gather structured and unstructured data, bring it into a common form, and add context to make a decision. Now, he says, it takes 10–15 minutes, and sometimes happens almost in real time.

The team that builds Codex itself, the code that lets it take actions, run commands, and use tools, describes a different effect. Adam Perry used to ask the agent to figure things out and dig around himself in parallel, because waiting was inefficient. Now he simply stays on a single task. Sayan Sisodia uses Codex not only for logs: the agent searches Slack for mentions of suspicious changes and complaints about similar symptoms elsewhere.

The 30-Second Threshold

Jeff Ngu states a simple rule: if the AI takes longer than 30 seconds to respond, he goes to Slack or messages a colleague. With Ultrafast, most of his requests fit within that threshold. Previously, he would ask a question, switch to something else, and come back having lost the thread of his reasoning. Now the exchange happens in real time.

For product and engineering teams, this is a useful benchmark. An agent has not only accuracy but also a human patience threshold. If answers take longer, people start duplicating the work by hand, and the benefit of automation evaporates. Speed determines whether a tool gets used at all.

What Changes Beyond Incidents

Ngu says the fast mode made him bolder with large refactorings: reworking existing code takes less time, and paying down technical debt is easier. In his view, code is cheap today, and refactoring that once cost a lot now happens almost instantly. This is one engineer's opinion rather than a measured result, but the direction is clear: when the cost of an iteration falls, people take on tasks they once postponed.

Perry maintains an ongoing dialogue with Codex, which he calls his "chief of staff." Each morning it reminds him of meetings and unanswered messages. Since the data is already gathered, follow-up questions do not require going back to the sources. Sisodia hands the agent routine browser work, such as filling out forms. He used to avoid this: the agent would "think" for five seconds before every click. With improvements in computer use and the faster mode, he now hands off such tasks without hesitation.

What to Take Away

First, for on-call duty and incident reviews, it is worth measuring not only the quality of the agent's hypotheses but also the time to the first useful answer. A threshold of tens of seconds is a reasonable target.

Second, parallel search across different sources, such as logs, chats, tickets, and security data, delivers its main gain specifically with fast responses. A slow agent forces people to search on their own.

Third, the cost needs to be counted soberly. OpenAI does not disclose what Ultrafast costs in tokens and infrastructure. Before bringing this approach to your own team, run your own pilot on real incidents and compare investigation time and request costs against the current process.

Story

Key Questions Answered

What It Looks Like During an Outage?

Aaron Fril, an OpenAI engineer, usually leads incidents. He creates a dedicated Slack channel and connects Codex: the agent is meant to monitor the situation, look for the cause, and answer questions as they come up. According to Fril, any outage has many possible links where something could have broken, and the main job of the people in the channel is to find the right needle in the right haystack. Ultrafast lets the agent search those haystacks simultaneously rather than one after another.

What are the key details on The 30-Second Threshold?

Jeff Ngu states a simple rule: if the AI takes longer than 30 seconds to respond, he goes to Slack or messages a colleague. With Ultrafast, most of his requests fit within that threshold. Previously, he would ask a question, switch to something else, and come back having lost the thread of his reasoning. Now the exchange happens in real time. For product and engineering teams, this is a useful benchmark. An agent has not only accuracy but also a human patience threshold.

What Changes Beyond Incidents?

Ngu says the fast mode made him bolder with large refactorings: reworking existing code takes less time, and paying down technical debt is easier. In his view, code is cheap today, and refactoring that once cost a lot now happens almost instantly. This is one engineer's opinion rather than a measured result, but the direction is clear: when the cost of an iteration falls, people take on tasks they once postponed.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…