Blog

If you count the incidents each provider posted on its own status page, OpenAI logged 40 in September 2026 and Anthropic logged 13, which looks like a clear win for Claude. But in July the same count ran the other way, with Anthropic at 55 and OpenAI at 36. We read both providers’ status pages and data feeds directly. The public data cannot separate two explanations: that each company logs incidents in its own way, or that one service really did fail more often in a given month. What we can show is that the count alone gives opposite answers in different months, and that the uptime figures each provider publishes sit within a fraction of a percentage point of each other, although they are calculated over different windows.

Key takeaways

1. OpenAI’s status page lists 40 incidents for September 2026 whether you group them by UTC or by UTC+5:30 time. Anthropic’s lists 13.
2. The lead flipped between August and September: Anthropic logged more in July (55 against 36) and August (25 against 18 or 19), then OpenAI logged about three times as many in September (40 against 13).
3. On a same-start-date window (incidents that began 14-30 September, UTC), OpenAI logged 23 incidents totalling 6,096 minutes and Anthropic 4 totalling 343 minutes. That excludes Anthropic’s Windows incident, which began on 10 September but was still open for about 988 minutes inside the window; counting that overlap, Anthropic’s total would be about 1,331 minutes. Either way, overlapping incidents add up, so none of these totals is downtime.
4. Impact labels are not comparable: OpenAI marked a 5-hour ChatGPT, Codex and API incident “minor” and a 56-minute Codex outage “critical”, while Anthropic marked 8 of its 13 September incidents “major” (62%, against 3 of 23, or 13%, at “major” or “critical” for OpenAI).
5. Published uptime figures are close but not like-for-like: on OpenAI’s page, ChatGPT 99.65%, Codex 99.95% and the APIs 99.96% (default July-October range); on Anthropic’s, claude.ai 99.59% for August and 99.69% for September.

Last reviewed 4 October 2026. We read OpenAI’s status history, its public incident feed and its RSS feed, and Anthropic’s status history, incident feed and uptime page, in a browser on that date. We did not test either service ourselves. Times in this article are UTC unless stated, because OpenAI’s page shows times in the viewer’s own time zone.

What does each status page count, month by month?

We counted incident entries per month on each provider’s own page. For OpenAI we counted the entries its history page (retrieved 2026-10-04) lists, and cross-checked with its RSS feed, which carries UTC timestamps. For Anthropic we used the “Show All” totals on its history page (retrieved 2026-10-04, except July, retrieved 2026-09-30).

Two things stand out. Anthropic logged more incidents than OpenAI in July (1.5 times as many) and August (about 1.4 times), and then OpenAI logged about 3.1 times as many in September. Each provider’s own total also moved a lot: Anthropic’s fell from 55 in July to 13 in September, and OpenAI’s more than doubled from August to September. Those swings could reflect real changes in how often each service failed, changes in how each company decides what to post, or both. The public pages do not let us separate those explanations, which is why we would not rank the two on this number.

The second point is smaller: OpenAI’s history page groups incidents by month in the viewer’s own time zone. An incident resolved on the evening of 31 August UTC counts as August in UTC but as September for a viewer in a UTC+5:30 zone, which is why our August count is 18 in one grouping and 19 in the other. The September total is 40 either way because the shift also runs the other way: an incident resolved on the evening of 30 September UTC counts as September in UTC but October at UTC+5:30. Anthropic’s page shows its incident times in UTC, and we did not re-group its counts.

Why a raw incident count is a poor reliability score

1. The two pages log different kinds of things. Both include non-availability items. OpenAI’s list for 14-30 September includes “Ads Manager login issues” (1 minute), “Delayed support responses” and an overbilling problem in the Agent API that stayed open for about 9 hours and 23 minutes with impact marked “none”. Anthropic’s September list includes “Issues with Google Play subscriptions” (33 minutes, impact “none”) and “Delays in credit purchases”. Neither list is only about whether chat works, so a single count mixes outages with billing, support and login issues.

2. Impact labels mean different things on each page. Across OpenAI’s 23 incidents started 14-30 September, 13 are labelled minor, 7 none, 2 major and 1 critical. Across Anthropic’s 13 September incidents, 8 are labelled major, 4 minor and 1 none. On OpenAI’s page, a 56-minute Codex outage on 25 September (22:58 to 23:54 UTC) is labelled critical, while a broad incident affecting ChatGPT, Codex and the API including the Agents API, from 29 September 17:52 to 23:14 UTC, about 5 hours 21 minutes, is labelled minor. We did not find a definition of either company’s impact labels on the pages we read, so we would not compare the labels across providers.

3. Duration and count tell different stories. The closest comparison we could build uses incidents that started from 14 September 02:46 UTC through 30 September, because that is as far back as OpenAI’s public JSON feed reaches. We call it a same-start-date window, not a like-for-like one, for a reason explained below the table. In that window:

Read that table carefully. Total minutes add up incidents that overlap and that affect different features, so it is a measure of how much was posted, not how long users were unable to work. About 45% of OpenAI’s total comes from one 45-hour incident affecting a single ChatGPT feature. The window also leaves something out on Anthropic’s side. Its biggest September incident, the roughly 99-hour Claude Cowork on Windows problem, started on 10 September and was still open until 14 September at 19:14 UTC, about 988 minutes after the window opens. Counting that overlap, Anthropic’s total in the window would be about 1,331 minutes instead of 343, which is why we do not call this like-for-like. For the whole month, Anthropic’s 13 incidents add up to about 114 hours, 99 of them from that one incident. We cannot give OpenAI’s equivalent, because its feed reaches back only to 14 September.

What do the published uptime numbers say?

Both providers also publish an uptime figure, which is a better starting point than a count, though still not a clean comparison. OpenAI’s status page (retrieved 2026-10-04) shows, for its default July to October 2026 range, ChatGPT at 99.65% (16 components), Codex at 99.95% and the APIs at 99.96%. Anthropic’s uptime page (retrieved 2026-10-04) shows claude.ai at 99.59% for August and 99.69% for September 2026, and lets you switch between its other components, which we did not read.

Those figures sit within about a tenth of a percentage point of each other, far closer than 40 against 13, although a count and a percentage measure different things. They are also not directly comparable: OpenAI’s is a range-wide figure for a group of components and its page does not say how the 16 ChatGPT components are combined, while Anthropic’s is per component per month. To put the percentages in context, 0.35% of a 30-day month is about 150 minutes and 0.1 of a percentage point is about 43 minutes, although OpenAI’s figure covers a range of roughly three months, not one. We would treat the two as being in the same range, not as a ranking.

What this does and does not tell you

• Supported: both providers logged incidents in every month we looked at, and September included a 5-hour 21-minute ChatGPT, Codex and API incident on 29 September (UTC) and a 2-hour 6-minute major Claude incident the same day. The Claude incident began 3 hours 31 minutes before the OpenAI one and had ended about 1 hour 25 minutes before it began, so they did not overlap. We make no claim that the two were related.
• Supported: counting incidents puts Anthropic ahead in July and August and OpenAI ahead in September, so one month’s count is not a reliability ranking.
• Not supported by what we read: any claim that one assistant is “more reliable” than the other. The pages measure different things, label severity differently, and expose different slices of history.

How to judge reliability for the tool you actually use

• Look at incidents touching your feature, not the total. If you use Codex or Claude Code, the Codex and Claude Code entries matter more than Ads Manager or billing.
• Read durations, not just counts. A single long incident can outweigh a dozen short ones.
• Check whether there is a write-up. OpenAI published a detailed one for its 3 September incident, which we covered in our check of the Azure explanation.
• Keep your own record. A status page reports aggregate monitoring, and as our earlier count of OpenAI’s incidents showed, your own experience can differ from what a page records.

What we could not verify

• We did not measure either service’s availability independently. Everything here comes from each company’s own status pages and feeds.
• OpenAI’s public JSON feed only reaches back to 14 September, so we could not compute exact durations for the first 17 of its 40 September incidents.
• We did not compare Google Gemini. We did not search beyond Google Cloud’s infrastructure status page for a status page covering the consumer Gemini app, so we make no claim about whether one exists.
• We read only claude.ai’s uptime on Anthropic’s uptime page, not its other components, and OpenAI’s uptime only as shown on its main page.
• Anthropic’s July figure of 55 was read on 30 September 2026. Its page now shows August onward, so we could not re-read it.

How we researched this

On 4 October 2026 we opened OpenAI’s status history and counted incident entries per month, then confirmed the counts against its RSS feed, which carries UTC timestamps, grouping by both UTC and UTC+5:30 time. We read OpenAI’s public incident feed for exact start and end times and impact labels, which covered 25 incidents from 14 September, 23 of them in September, and its RSS feed for UTC resolution times (both retrieved 2026-10-04). For Anthropic we read its history page and its public incident feed (retrieved 2026-10-04), which lists its 50 most recent incidents with exact timestamps and impact labels, and its uptime page. Where a figure is our own calculation, such as totals and medians, we say so. Where we label something as a limit, it is a limit of what the pages expose, not a claim about the services.

Frequently asked questions