When something you depend on breaks, the status page is where the company says so. We read a full year of incidents, 1 October 2025 to 30 September 2026, from the public status pages of 47 software services: AI assistants, developer tools, cloud hosts, work apps. Then we counted what they posted, how long incidents stayed open, and what came after.
- 3,263 unplanned incidents in a year. The typical service posted 3.3 a month; 17 of 45 posted at least one a week. A busy page isn't a worse service: what counts as "an incident" is each company's call.
- The two big AI labs post the most. Claude's page posted 431 unplanned incidents (35.4 a month); OpenAI's posted 29.6 a month over the 72 days of the window its page still shows. Over those same 72 days the two were close: 62 for Claude, 71 for OpenAI.
- Postmortems are rare. 27 of the 45 full-year pages never published one in the year, and only 297 of the 3,179 incidents on them (9%) got one. Smartsheet, Snowflake, Box, Trello and Linode are the exceptions; OpenAI's page published a write-up for 12 of its 71 (17%) in the 72 days it shows.
- "Major" means different things on different pages. HubSpot rated all 55 of its incidents "none"; Elastic rated 72 of its 84 "major".
- Most incidents close within two hours (median 1.9 h), but 1 in 9 (11.2%) stayed open for more than a day.
It's great to see teams trying to tell their users what's going on. I check Claude's status page whenever my Claude agent has problems, and it's usually right about which model is affected: 235 of its 431 incidents here name one in the title. It also tells you whether a fix is under way, or whether they've only had reports of degraded performance so far.
| What we measured (1 Oct 2025 to 30 Sep 2026) | Result |
|---|---|
| Status pages with a year of history | 45 (+ OpenAI's 72 days) |
| Unplanned incidents posted | 3,263 |
| Planned maintenance posted (counted apart) | 614 |
| Unplanned incidents per month, the median service | 3.3 |
| Services posting at least one a week | 17 of 45 |
| Median time from posted to resolved | 1.9 hours |
| Open for more than a day | 11.2% |
| Pages that published at least one postmortem | 18 of 45 (and OpenAI's) |
| Incidents whose start was set earlier than the post | 4 of 3,263 |
How often services post
The busiest page in our set is Scaleway's (529 unplanned incidents, 43.5 a month). It posts one for each product and region, like "[Kafka] - [fr-par] Control plane partial outage". Coinbase (28.7 a month) posts one per coin whose transfers are slow: "XLM - Delayed Sends & Receives". Gemini (the crypto exchange, 15.0 a month) is next.
At the other end, Jotform posted 3 incidents in the whole year, Python's infrastructure page 6, and Hotjar, RubyGems.org and Typeform 7 each.
Neither end tells you which service failed you more. A company that posts every slow coin transfer looks busy; one that posts only when everything is down looks calm. Read the counts as a measure of how much each page tells you.
| Group (the median service; pages with 60+ days of history) | Services | Incidents a month | Median hours open | Rated major or critical |
|---|---|---|---|---|
| Developer tools and hosting | 17 | 3.5 | 2.2 | 30% |
| Storage, web and payments | 10 | 3.7 | 1.7 | 19% |
| Work and chat apps | 16 | 2.1 | 1.8 | 40% |
| AI (Claude, OpenAI, Runway) | 3 | 29.6 | 1.5 | 35%* |
* Claude and Runway only: OpenAI's page rates components on another scale (below), so it is left out of that column.
The AI services
The two biggest AI labs run the busiest pages in the set after Scaleway. Claude's page posted 431 unplanned incidents in the year, from "Elevated errors on Claude Opus 4.7" to "Claude Desktop is unavailable". OpenAI's page only shows its last few months: from 21 July 2026 it posted 71 to the end of September.
Over that same span, 21 July to 30 September 2026, the two are close: Claude 62, OpenAI 71. Both closed the typical incident within two hours (median 1.06 h for Claude, 1.85 h for OpenAI). Runway, the third AI service in our set, posted 4 a month.
Don't compare their severity. In that span Claude rated 39% of its incidents major or critical and OpenAI 8%, but the two use different status page software with different scales: Claude's page (Atlassian Statuspage) rates each incident; OpenAI's (incident.io) marks each affected component. OpenAI's "Chatgpt.com is down - all signups and logins are down as of right now" on 20 August 2026 (UTC) marked one component as "degraded performance".
Postmortems: the exception, not the rule
A postmortem is the write-up after an incident: what broke, why, and what changes. Statuspage has a field for it; incident.io pages attach a write-up to the incident.
27 of the 45 full-year pages never used it. Across the 3,179 incidents on those pages, 297 got a postmortem (9%). A few companies do it as a habit: Smartsheet published one for 34 of its 35 incidents, Snowflake for 58 of 63, Box for 42 of 50, Trello for 8 of 12 and Linode for 73 of 110. Claude's page published 9 for its 431 incidents, all before 21 July. OpenAI's page published a write-up for 12 of its 71 from 21 July, the 20 August ChatGPT outage among them; over that same span Claude's page published none.
Plenty of companies publish postmortems elsewhere, on a blog or by email to customers. But the status page is where people look while something is broken, and on most of them the story ends at "Resolved".
"Major" isn't a unit
Status pages let each company rate its own incidents: none, minor, major or critical. They use the scale very differently.
- HubSpot rated all 55 of its incidents "none", including four whose titles begin "HubSpot is experiencing issues".
- Coinbase rated 218 of its 349 "none" and 4 major or critical.
- Elastic rated 72 of its 84 "major", including 5 minutes of "Intermittent 404s when accessing elastic.co webpage".
- Box rated 43 of its 50 major or critical.
None of this says whose incidents were worse. It says the colour of the banner depends on who paints it. To compare two services, read what broke and for how long. We found the same thing on crypto exchanges' status pages.
How long incidents stay open
The median unplanned incident was resolved 1.9 hours after it was posted. The long tail is long: 364 of 3,260 resolved incidents (11.2%) stayed open more than a day, and 112 more than a week. On Scaleway's page one incident in ten stayed open for over a week (90th percentile 176 hours), and on Gemini's nearly as many (9%, 157 hours): both post one incident per product or per coin, and a single product or coin can stay degraded for days.
What a status page can't tell you
When the problem really started. A status page lets a company date an incident's start earlier than the moment it posts. In 3,263 incidents that happened four times: once on Claude's page ("iOS Subscription Access Issues", posted with a start 19.6 hours earlier) and three times on OpenAI's, where an affected component's impact starts 1.9 to 9.4 hours before the first post. Everywhere else the clock starts at the post. So every duration on a status page, and every one in this article, measures from when the company said something, not from when users were first affected.
What it didn't post. A status page is self-reported. An outage nobody posted doesn't exist on it, and some pages don't keep their history: OpenAI's goes back about three months, Atlassian's main page posted nothing in the year, and we couldn't read a year back on 10 of the 58 pages we started with (Cloudflare's, for one).
Whether it's down for you. A service can be "operational" on its page and failing from your network, your region or your account.
That last one is what an independent check is for. UpWatchr, our free Windows app, checks the sites and services you depend on from your own PC, on the interval you set, and alerts you at its next check when one stops answering, whatever its status page says. For a one-off look, the HTTP status checker shows what a URL answers right now.
How we tested
| Services | The 58 software services we found with a Statuspage-format status page on 8 October 2026 (from 131 checked), in four groups: AI, developer tools and hosting, storage/web/payments, work and chat apps |
| Period | Incidents posted 1 October 2025 to 30 September 2026 (UTC). OpenAI: 21 July to 30 September 2026, the 72 days of the window its page still shows. Fly.io's history starts 2 September (29 days of the window) and is left out of the comparisons |
| Data | Each page's public incident history, then each incident's own record: created and resolved times, impact rating, updates, postmortem (incident.io: the write-up). Read on 23 October 2026, one request a second per site, with a user agent naming us |
| Unplanned vs planned | Scheduled maintenance counted apart; everything else is unplanned |
| Duration | From the time the incident was posted to the time it was marked resolved, as the page stamped them |
| Severity | The page's own rating. OpenAI's page rates components, not incidents: we took its worst component status (degraded performance = minor, partial outage = major, full outage = critical), and don't compare it with the others |
| Not covered | 10 of the 58: Cloudflare, Canva, Loom, Zapier, Cohere, ClickUp, ElevenLabs, Miro, Mixpanel and Twitch (history not readable back a year). Atlassian's main page was read and had no incident in the year |
| Caveat | Self-reported: an incident a company didn't post isn't here, and its timing is the company's own |
Download the data: pages_history.csv (one row per service) · incidents_history.csv (all 3,877 incidents: service, title, impact, created, resolved, hours, postmortem) · the CSV behind each chart : u3-incidents-per-month.csv, u3-postmortems.csv.
FAQ
How often do SaaS services have outages? By what they post, the typical service in our set posted 3.3 unplanned incidents a month in the year to September 2026, and 17 of 45 posted at least one a week. Many are partial: one region, one product or one coin, as their titles show.
How long does a typical outage last? The median incident was resolved 1.9 hours after it was posted; 11.2% stayed open for more than a day.
Does ChatGPT go down more often than Claude? Over the same 72 days (21 July to 30 September 2026) OpenAI's status page posted 71 unplanned incidents and Claude's 62. That's close, and both are counts of what each company chose to post. Their severity ratings use different scales and can't be compared.
What is a postmortem? A write-up after an incident explaining what broke, why, and what will change. On status pages they are rare: 27 of the 45 full-year pages published none in the year.
Can I trust a status page? As a record of what the company says, yes. As a measurement, no: it shows what was posted and when, not what you saw. Durations start when the incident was posted, and only 4 of 3,263 incidents had an earlier start set. To know when a service is down for you, check it yourself.
