OpenAI Uses Codex Ultrafast Mode to Cut Incident Investigation Time From Hours to Minutes
OpenAI engineers are using a high-speed variant of Codex to slash incident response timelines, reducing security triage from 60–120 minutes to 10–15 minutes.
Last verified:
Parallel Hypothesis Testing Replaces Sequential Investigation
According to OpenAI’s blog, incident response engineers at the company are using Codex Ultrafast—a high-speed inference mode—to investigate multiple potential causes of system failures simultaneously rather than pursuing leads sequentially. Previously, when infrastructure or service issues emerged, engineers faced a choice: wait minutes for Codex to analyze logs and configuration data, or manually hunt for root causes while an outage compounded. Ultrafast eliminates that trade-off by returning structured answers within 30 seconds, allowing teams to evaluate competing hypotheses without losing focus.
From Hour-Long Triage to 10-Minute Resolution Windows
The operational speedup is most visible in security incident investigation. According to OpenAI, AJ Trbojevic, a Member of the Technical Staff and Datacenter Security Architect, noted that pulling together structured and unstructured data—logs, configuration snapshots, alert timelines—and normalizing it to make a containment decision historically consumed 60–120 minutes. With Ultrafast, the same investigation now completes in 10–15 minutes, and in some cases approaches near-real-time resolution.
OpenAI’s blog post highlights how this acceleration changes team behavior during incidents. Aaron Friel, a Member of the Technical Staff who typically leads incident response, uses Codex Ultrafast to monitor multiple hypothesis paths at once. Rather than testing one potential cause, waiting for results, then moving to the next, Friel can interrogate the logs and system state across multiple dimensions in parallel, collapsing the investigation timeline.
Eliminating Context Switching During Active Incidents
The sub-30-second response threshold addresses a secondary but equally important friction point: context-switching overhead. According to OpenAI’s post, Jeff Niu, another Member of the Technical Staff, previously would ask Codex a diagnostic question, then switch to Slack or another tool while waiting for the reply, often losing his train of thought by the time the answer arrived. Ultrafast keeps the conversation in real time, maintaining focus on the incident itself rather than auxiliary tasks.
Adam Perry and Sayan Sisodiya, both on OpenAI’s core agent team, report a similar shift in workflow. Before Ultrafast, they would start a Codex investigation in parallel with their own manual review, hedging against slow AI responses. Now, per OpenAI’s account, they can rely on Codex to be fast enough that splitting attention between manual and AI-assisted investigation is no longer necessary. Sisodiya also leverages Codex to cross-reference Slack messages during security triage—searching for discussion threads mentioning recent code changes or similar symptoms elsewhere in the organization.
Why This Matters
This case study demonstrates a concrete efficiency gain in a high-stakes operational domain. OpenAI’s incident response SLAs—both for availability and security—now have tighter time budgets for investigation. Teams can commit to sub-15-minute triage windows for security incidents because the bottleneck of data assembly and normalization has moved from serial manual work to parallel AI-assisted queries. The shift also reveals how inference latency, not just model capability, shapes adoption of AI-assisted workflows. Codex was available to OpenAI’s teams before Ultrafast; the speed increase itself is the enabling factor, suggesting that other organizations operating their own high-latency LLM infrastructure may see similar gains if they prioritize sub-30-second inference SLAs for synchronous incident response tasks.
Frequently Asked Questions
What is Ultrafast mode in Codex?
Ultrafast is a high-speed inference variant within OpenAI's Codex that returns answers in under 30 seconds, enabling real-time incident investigation rather than sequential tool queries.
How much faster is the investigation process now?
According to OpenAI, security incident investigations that previously required 60–120 minutes now complete in 10–15 minutes, with near-real-time responses in some cases.
Is this a new product or an internal tool?
This is an internal use case OpenAI is sharing as an example of how its own frontier AI models support internal workflows; it is not a separately announced product.