Codex Ultrafast cut security investigations at OpenAI from an hour to 15 minutes. Customers pay six times the price

OpenAI has published accounts from its own engineers about how the Ultrafast mode in Codex changes their work during outages. The biggest gain they describe is attention: they no longer investigate in parallel with the agent, because the answer arrives before they would have found anything themselves.

Ultrafast searches logs and Slack before the engineer loses the thread

Ultrafast is a high-speed mode in Codex, and OpenAI says it has become the default for its engineers during incidents. The post belongs to a series in which the company shows its own internal deployments. Aaron Friel, who usually runs point on incidents, opens a Slack channel and has Codex monitor, investigate and answer questions as they come in. Thanks to the speed, he says, he can search all the haystacks at the same time.

The most concrete number comes from AJ Trbojevic, a datacenter security architect. Pulling structured and unstructured data together, normalizing it and preparing context for a decision used to take an hour or two. Now it takes 10 to 15 minutes, he says, and sometimes it is close to real time. Sayan Sisodiya, who works on the team building Codex’s agent core, also has the agent dig through Slack for mentions of suspect changes.

Jeff Niu has a simple rule: if an AI answer takes more than 30 seconds, he asks a coworker instead. With Ultrafast, he says, most of his work stays under that threshold, which makes him more willing to take on large refactors. Adam Perry keeps one long-running Codex thread he calls his chief of staff. Every morning it lists his meetings and the messages he forgot to answer.

Speed changes what is worth handing to an agent

OpenAI launched Ultrafast for GPT-6.1 Sol on October 8. According to the developer announcement, it is up to 8x faster than standard Sol. In the API it costs $12 per million input tokens and $60 per million output tokens, six times the standard price. In Codex and ChatGPT Work it is available on the Pro 500 plan, eligible usage-based Enterprise accounts and credit-based Edu plans, and Enterprise admins have to switch it on. OpenAI also lists data residency support in the US and the EU, so European companies can consider it for data that must stay in Europe.

What the engineers say between the lines is more interesting than the numbers. Adam Perry says he used to give Codex a task and then chase the same lead himself, because waiting on the agent was not worth it. A slow agent did not save work, it doubled it. Only when the answer comes within half a minute do people stop investigating in parallel.

That leaves team leads with a practical question. Paying six times more makes sense where waiting costs more than tokens, typically during an outage. For everyday coding, when nothing is on fire, it is an expensive luxury.

The witnesses are the vendor’s own staff, who do not pay for tokens

Every account comes from OpenAI employees who build the tool or use it without budget worries. The post does not say how many incidents were measured, how often Codex pointed the wrong way or what an investigation cost. The only hard number, 10 to 15 minutes instead of an hour or two, is one person’s estimate.

A fast wrong answer is more dangerous during an incident than a slow one. Anyone who lets an agent into logs, Slack and a browser also has to settle permissions: what the agent may read, what it may change and who checks its work. The post says nothing about those safeguards.

Numbers from companies paying for Ultrafast themselves will decide

The first signal will be case studies outside OpenAI with incident counts and time to resolution, ideally from companies that pay for the fast mode out of their own budget. The second is whether Ultrafast reaches plans cheaper than Pro 500.

It is also worth watching how OpenAI describes agent permissions during incidents. Companies that want Codex reading production logs will need a clear audit of what the agent saw and did before they need speed.

Lilith's verdict

The engineer who used to hunt for the bug alongside the agent, because waiting for it was pointless, now calmly reads its answers. What OpenAI is really selling is half a minute of focus, billed at six times the token price.