The notice appeared at 8:10 a.m. Coordinated Universal Time. “We are investigating the issue for the listed services,” OpenAI wrote on its status page on Tuesday, listing ChatGPT and Codex as affected.
Eight minutes later the company had narrowed the fault to a name. The degraded performance, it said, was in GPT-6.1 Sol Codex and Work mode, the two products OpenAI sells to developers and companies that use its models to write and run software. A further update at 8:22 said the investigation was continuing. At 9:30 the company said it had applied a mitigation and was monitoring the recovery. That was the last entry it posted. By the end of the trading day in New York there was no root cause, no explanation of what the mitigation covered, and no notice that the incident had closed.
For someone chatting with a bot, an afternoon of slower replies is an annoyance. For the buyers of these two products it is a cost. Codex works inside a developer’s editor, terminal and web console, and its value comes from finishing tasks unattended. Work mode is sold to companies that hand routine jobs to a model and expect the job done. Both are billed by subscription or by usage, which means a model that answers late consumes the same budget for less work and interrupts the pipelines a company has built around it.
Degraded performance is the hardest kind of failure to see from outside. A service that is down announces itself: requests fail, monitoring turns red, users complain within minutes. A service that is slow, inconsistent or quietly worse at following instructions keeps returning successful responses while users retry prompts and rerun jobs. UptimeRobot, which tracks the page, made the point after the incident: requests can keep succeeding while answers arrive slower, less consistently or partly wrong, and a simple up-or-down check will not catch any of that.
OpenAI’s own figures describe a service that is available almost all of the time. The status page reports 99.52% uptime for ChatGPT from July through October, 99.87% for ChatGPT Work, 99.89% for Codex in ChatGPT Desktop and 99.95% for Codex. The numbers are accurate and, for a buyer, incomplete. A quarterly figure of 99.5% still leaves room for hours in which the product accepted work and did it badly.
The incident is the latest in a sequence. On Sept. 29 OpenAI opened a notice for elevated errors across ChatGPT, Codex and its API, including the Agents API, affecting Codex Web, the Codex API, the CLI and an editor extension. It declared full recovery on Oct. 6 and said a root cause analysis would be published within five business days. A separate listing described delayed responses, turn failures and reduced proactivity in Codex and Work mode. The company, like its rivals, publishes the raw material of these events and says little about their causes.
For three years the labs competed on capability: longer context windows, stronger reasoning, a lower price per token. The buyers have changed, and so have their questions. Banks, insurers and software companies are now wiring agents into support desks, procurement and engineering, where a task either completes or it does not. Executives who once asked what a model could do now ask what happens when it stops.
Analysts said enterprise buyers have begun to write response times and fallback requirements into contracts, terms that are easier to agree than to guarantee at the scale these systems run. The mismatch is structural. Models are served from data centers that draw power from grids, and each layer can fail in ways that show up as slowness rather than silence.
Reliability has a second cost that a status page cannot show. Agentic products are chained: a single request inside a coding assistant may spawn a dozen further calls, each of which reads a repository or runs a test. A slowdown at any link lengthens the chain, and the customer pays for the whole chain. Failures that would be invisible in a chat session become expensive inside a workflow.
What the vendors can do about it is limited and familiar. They route requests to smaller models when the largest one falters, roll back changes and buy time. Those techniques work, and each one trades quality for continuity, a trade the customer is rarely told about. Rivals face the same problem. Every large provider runs a status page, and all of them have posted a sentence close to this week’s.
None of this is evidence that OpenAI’s systems are unusually fragile. What it shows is that the company’s most valuable customers have arranged their work around the assumption that a model call returns promptly, and that assumption is covered by nothing the company publishes. A model that is late is not a broken model. It is a broken contract, and the buyer discovers it the way a driver discovers a toll increase: on the bill.
Two things are worth watching. The first is whether the root cause analysis promised for the September incident arrives, and whether this week’s notice closes with an explanation rather than a mitigation. The second is whether buyers keep signing contracts that trade price for availability. Neither would settle the argument about what these models can do. Together they would show that the binding limit has moved from capability to keeping the lights on.
