OpenAI Logged 12 Incidents In 10 Days. Then Codex Went Down.
A critical Codex outage on September 25, about 8,000 Downdetector reports, and paying users who say their usage started draining right after.
Server racks in a data center. Photo: BalticServers.com, CC BY-SA 3.0, via Wikimedia Commons.
You don't need Downdetector to see how rough the last two weeks have been for ChatGPT. OpenAI's own status page tells the story. Between September 16 and September 25, 2026 it logged 12 separate incidents.
The run of incidents
Here's what the status page shows, in Pacific time. ChatGPT Work had elevated errors on September 16 and a major incident on September 17, when some users couldn't start or run Work tasks. API models threw elevated errors the same afternoon. On September 18 single sign on broke for some business customers, support replies were delayed, and OpenAI disclosed it had been overcharging for hosted containers in its Agent API and said it was preparing refunds. Plus and Pro users hit elevated errors on September 22, again for ChatGPT Work that morning, and again overnight into September 23. That afternoon mobile users couldn't see Work mode or the model picker. On September 24 GPT 6 Astra Pro had elevated errors.
September 25: Codex goes down
At 3:58 PM Pacific on Friday, September 25, OpenAI posted that "users are experiencing a codex outage." It was rated critical. At 4:19 PM the only advice was a workaround: "Login via API key will unblock access at this time." The status page marked everything recovered at 4:54 PM, 56 minutes after the first post.
GV Wire, tracking Downdetector, counted more than 2,700 problem reports at 3:55 PM, more than 4,000 a few minutes later and nearly 8,000 by 4:25 PM. Once it was back, Tibo, who runs Codex, said on X that usage limits would be reset "for all paid users across codex and ChatGPT work." That didn't land right away either. More than 12 hours later many users were still reporting unrefreshed limits, and the reset didn't finish until around 11:17 AM Pacific on September 26.
Then the limits started melting
The day after, a Plus user posting as hirohamada on OpenAI's community forum said a single test message after the recovery went badly wrong:
A second one line test, they wrote, took more than 30% of the five hour limit and 9% of the week. "At this rate, I cannot use Work for my actual work." On September 27 a Pro 5x subscriber, Tchatter, said the same thing was happening to them, including with purchased credits: "Normal or relatively short sessions appear to be consuming a disproportionately large percentage of my usage allowance." OpenAI's support account replied by asking which plan the first user was on. As of this writing the thread has no fix.
Why this belongs here
People pay OpenAI for predictable access. A dozen incidents in ten days, a critical outage at the end of the work week, and a quota meter nobody can explain afterward is the opposite of that. We've been collecting the complaints on our firsthand accounts page, and this week added a lot of them.
Sources: OpenAI status page incident history, GV Wire, Pasquale Pillitteri, explainx.ai,OpenAI Community forum thread.