At 12:25 Thursday afternoon, six of Hacker News's top 25 stories were about trouble at Grok, ChatGPT, Claude or Codex, including four of the top twelve.1
The front page carried separate discussions for Grok, ChatGPT, Claude and Codex, a report about all three providers and an Ask HN thread about why the failures arrived together. Hacker News is not a survey of developers. The comments show what the interruption looked like among people who had already put these models inside daily software work.
three provider records overlapped for 93 minutes
Anthropic's incident record begins at 9:26 a.m. Eastern, xAI's begins at 9:30, and OpenAI posted its first update at 10:43. Anthropic reported impact ending at 12:16 and marked its incident resolved at 12:23; OpenAI marked its incident resolved at 12:55, followed by xAI at 1:07.2
All three records were active from 10:43 until 12:16. The chart compares the providers' public status records rather than the exact start and end of every user's trouble. The providers have not published a common cause.
the comments were about interrupted work
One programmer with twenty years of experience wrote that he was seriously concerned he might not be able to code anymore. Another said models had become part of his thinking and that he had not written a complete function in two months. A third used the interruption to answer email and work through architecture.3
Other accounts showed why a simple “up” or “down” label did not fit. Old Codex sessions continued while new sessions returned 404s. A paid ChatGPT account worked while logged out but failed on logged-in devices. One user reached the models through Amazon Bedrock while the providers' direct services failed.4
Several people moved to Gemini, OpenRouter, Bedrock or local models. Those routes gave some users another way to continue. Three providers had reported trouble within 77 minutes, and people saw different results by session, account and route.
McQueen had access and still had to validate it
McQueen Analytics uses several hosted frontier models and keeps open-weight models on our own hardware. Agents help us compare survey questions across waves, rerun calculations against the source and build software around methods shaped across more than twenty years of studies.5
We retained ChatGPT access Thursday: the page stayed open and replies kept arriving, but some output became visibly abnormal. I checked the work against its sources before using it.
Another model was available, but a second answer would not validate the first one. In research, a 404 is easier to handle than abnormal output inside a finished-looking answer. The 404 stops the work; the answer can travel into a questionnaire, calculation or client decision unless somebody checks it.
the dependence appeared all at once
When Cloudflare struggles, a long list of unrelated websites can struggle with it, making a shared dependency visible. Thursday's incidents disrupted coding, research and operations during the same morning. The comparison is reach, not cause; the provider records do not identify a shared dependency.
The 12:25 snapshot contained six Hacker News front-page stories about people trying to resume existing work. By 1:07 p.m., all three providers had marked their incidents resolved.
Incident records and discussions
- OpenAI: Elevated errors across ChatGPT and Codex
- Anthropic: Elevated errors for multiple models
- xAI: Grok models outage
- Grok outage discussion
- ChatGPT outage discussion
- Claude outage discussion
- Codex outage discussion
- Ask HN about the simultaneous timing
Source notes
- The front-page image is Carl McQueen's browser screenshot from about 12:25 p.m. Eastern on September 3. The six AI-service stories occupied positions 2, 6, 9, 12, 18 and 25 at that moment. The pages were the Grok outage, ChatGPT 404, Claude errors, multi-provider report, Codex outage and Ask HN discussions. Rankings and comment counts changed after capture.
- The provider records establish three incidents on the same morning. Anthropic began its record at 9:26 a.m. Eastern, reported impact ending at 12:16 p.m. and marked the incident resolved at 12:23. xAI began its Grok Web record at 9:30 a.m. and marked it resolved at 1:07 p.m. OpenAI posted its first update at 10:43 a.m. and marked the incident resolved at 12:55 p.m. The records do not identify a shared cause.
- The developer accounts come from direct Hacker News comments about concern about coding ability after twenty years, models becoming part of a developer's thinking and switching to email and architecture during the outage. They are examples of heavy use in a self-selected technical community, not estimates for all developers.
- The uneven-access examples are individual reports in the Codex, ChatGPT and Claude discussions. Direct comments include old sessions working while new ones failed, logged-out access working while logged-in access failed, Bedrock routes working while direct services failed and provider, OpenRouter and local-model fallbacks. These reports establish what those users saw, not a provider diagnosis.
- The description of McQueen Analytics' use follows the earlier RadiationBox articles We are using AI to do better research and Open-weight AI rules should preserve access for smaller organizations. It describes our own work rather than a claim about general business adoption.