Claude Status Page History, July to October 2026: We Sorted All 50 Incidents by Model and Product

Anthropic’s status page listed 50 incidents between 27 July and 7 October 2026, and they were not all the same kind of problem. Sorting them by what each title names shows a split by duration. Incidents whose titles say “multiple”, “many” or “all” models, such as “Elevated errors for multiple models”, stayed open for a median of 167 minutes, and 7 of 11 passed two hours. Incidents whose titles name one model, such as “Elevated errors for Claude Opus 5.5”, had a median of 30 minutes, and incidents naming two or three models had a median of 80. We built the dataset from Anthropic’s public feed, published it as a CSV, and use it here to answer the questions people search: which Claude models and products show up, how long incidents last, when they start, and whether Anthropic says why.

Key takeaways

1. 50 incidents started between 27 July and 7 October 2026, 49 of them resolved. The median stayed open for 63 minutes, the mean was 210 (90 without the 99-hour Windows incident) and the longest 5,961 minutes. 12 stayed open for more than two hours and 3 for more than six.
2. By title: incidents saying multiple, many or all models (11) had a median of 167 minutes and 7 passed two hours. Incidents naming one model (11) had a median of 30 and two or three models (7) a median of 80, and none of those 18 stayed open on the page for more than two hours.
3. 22 of the 50 (44%) were labelled major or critical by Anthropic. In 3 of the 50, an update attributes the problem to a third party. Most others describe what failed but not why, and Anthropic’s two engineering postmortems that we found cover response quality, not availability.
4. 49 of the 50 started on a weekday, which is 71% of days, with none on a Saturday. 22 (44%) started between 13:00 and 18:59 UTC, which is 25% of the hours. We cannot say why.

Last reviewed 11 October 2026. Data comes from Anthropic’s incident feed, retrieved on 10 October. The feed holds its 50 most recent incidents, the first starting on 27 July, and shows each incident’s final title. All times are UTC. The categories are our own reading of each incident’s title, and the CSV and rules are linked in the method section.

What did we measure, and how?

For each incident in the feed we took the start as the time of its creation, the end as the time of its first “resolved” update, the severity as Anthropic’s own incident impact label (none, minor, major or critical), and the components as any component an update set to a non-operational status. One incident, “Elevated errors on platform.claude.com” from 7 October, had no resolved update when we read the feed (it was in a monitoring stage), so duration figures cover the other 49. We then sorted incidents into five groups by their titles: those that say multiple, many or all models; those that name exactly one model; those that name two or three; those about billing and credits; and the rest, which are about a Claude product or service. That sorting is our judgement from the wording of titles that Anthropic’s staff write, and the rules are in the script.

Measure (50 incidents, 27 Jul to 7 Oct 2026)	Result
Resolved / still open when read	49 / 1
Median minutes open (resolved)	63
Mean / 90th percentile / longest	210 (90 without the longest) / 257 / 5,961 minutes
Open for more than 1 hour / 2 hours / 6 hours	26 / 12 / 3 of 49
Impact labels	Minor 25, major 20, critical 2, none 3
Incidents by month of start	July (from 27th) 6, August 25, September 13, October (to 7th) 6
Started on a weekday	49 of 50 (Monday 11, Tuesday 11, Wednesday 10, Thursday 10, Friday 7, Sunday 1)
Started 13:00 to 18:59 UTC	22 of 50 (44%)
An update attributes the problem to a third party	3 of 50

The 90th percentile is the 45th of the 49 sorted durations. The longest incident, 5,961 minutes, is the Windows-update problem that left Claude Cowork on Windows unable to run local commands, which we covered in our article on it. The other two over six hours are “Degraded performance of multiple models” on 5 August (429 minutes) and “Delayed credits on the Claude Platform” on 1 October (376 minutes).

Do incidents that name several models last longer?

What the title names	Incidents	Median minutes open	Longest	Major or critical	Open more than 2 hours
One model	11	30	63	4	0 of 11
Two or three models	7	80	95	4	0 of 7
Multiple, many or all models	11	167	429	7	7 of 11
A Claude product or service	17 (16 resolved)	67.5	5,961	7	4 of 16
Billing, credits or subscriptions	4	75.5	376	0	1 of 4

Duration rises with how many models the title names. One model: median 30 minutes. Two or three: 80. Multiple, many or all: 167, with 7 of 11 lasting more than two hours while the other four closed in 26 to 98 minutes. Among the 18 incidents that name one to three models, none stayed open on the status page for more than two hours, and the longest was 95 minutes. The two incidents labelled critical, on 29 July (“Elevated errors across all models”, 167 minutes) and 16 August (“Service disruption on Claude services”, 36 minutes, filed here under products), were the widest in wording.

Three cautions apply. First, the groups are small, 7 to 11 incidents each, and the split rests on titles, which Anthropic’s staff write and can edit: the 16 earliest incidents in the feed were all last modified in a single batch on 14 August, so their final titles may differ from what was first posted. Second, if the two “Service disruption on Claude services” incidents (88 and 36 minutes) are counted as wide-scope, the several-models group has 13 incidents and 7 of 13, or 54%, passed two hours. Third, the page time is not always the impact time. Twelve resolved updates state their own impact window, and in several it differs from the page: the 14 August “Elevated errors for Claude Fable 5” incident was open on the page for 30 minutes, but its resolved note gives about four hours of impact, 20:00 to 00:11 UTC. With the stated windows used where they exist, that incident would pass two hours, so a claim that no incident naming one to three models ran past two hours does not hold. We expect the gap between the groups to survive, since most of the several-models incidents are long on the page too, but we have not recomputed it on stated windows.

We do not know why multi-model incidents last longer. A title that says “multiple models” may describe a wider fault, such as shared infrastructure, but Anthropic gives no cause for any of these 11, and a title that names one model may simply be where the problem was first noticed. What the data supports is that when Anthropic describes an incident as spanning several models, it has tended to stay open for hours.

Which models appear in incident titles?

Model family named in a title	Incidents naming it	First and last	Note
Sonnet (Sonnet 5)	8	31 Jul to 3 Sep	Not named in a title since 3 September
Opus (4.8, 5 and 5.5)	7	27 Jul to 6 Oct	Opus 5 five times to 19 August, Opus 4.8 once, Opus 5.5 once (6 Oct, 18 minutes)
Fable (5 and 5.1)	5	13 Aug to 5 Oct	Three of the five name Fable 5.1, with Mythos 5.1
Mythos (5 and 5.1)	4	13 Aug to 5 Oct	Always named together with a Fable model
Haiku (4.5)	2	27 Jul and 19 Aug	Both alongside Opus 5

The table counts titles, so an incident that names two families counts for each. Sonnet 5 appears in titles from 31 July to 3 September and not again after that, Fable and Mythos 5.1 appear from 11 September (lasting 16, 17 and 86 minutes), and Opus 5.5 appears once, on 6 October (18 minutes, labelled major). We do not know when each model was released or what changed between them, and which model a title names reflects which one Anthropic chose to name, so we draw no conclusion about the models’ relative reliability.

Which products are affected?

Counting components that an update set to a non-operational status, claude.ai was touched by 38 incidents, Claude Code by 35, Claude Cowork by 35, the Claude API by 34 and Claude Console (platform.claude.com) by 10. An incident that marks several components counts for each, which is why the first four are so close, and many of the same incidents appear in all of them. These are counts of incidents, not of time: our rebuild of both providers’ uptime found claude.ai, for example, spent about 1.2% of 1 August to 9 October in a degraded-only state, with a further 833 minutes at partial outage and 36 at major outage.

The 17 product incidents (median 67.5 minutes) are a mixed group: sign-in problems on 24 August, Office 365 degradation on 1 September, Google connectors on 20 August, “Service disruption on Claude services” on 14 and 16 August, a problem reaching the status page itself on 14 August, and the platform.claude.com errors on 7 October. The four billing incidents are credit purchase delays on 1 September and 1 October, a Google Play subscription problem on 16 September and incorrect spend-limit pauses on 7 October, a zero-minute entry that was posted after the fact.

When do incidents start?

49 of the 50 incidents started on a weekday, which is 71% of calendar days, and none started on a Saturday. By hour, 22 (44%) started between 13:00 and 18:59 UTC, which is 9 a.m. to 3 p.m. US Eastern (EDT, UTC minus 4) and 25% of the hours in a day. We picked that window after seeing the data, though the weekday skew is much larger than chance would give. The page’s start times follow when staff post as well as when problems begin, so we cannot say whether this reflects changes shipped during US working hours, user load or posting habits. For OpenAI’s incidents we saw a similar weekday skew, and there the write-ups did not point to working-hours changes, as we covered in our dataset of OpenAI’s 111 incidents.

Does Anthropic say what went wrong?

Rarely, on its status page. In 3 of the 50 incidents an update attributes the problem to a third party: the Windows update that broke Cowork on Windows (a fix from Microsoft closed it), “an issue with an upstream cloud provider” on 28 August, which it did not name, and a Google Play problem that Google had confirmed. Other updates describe what failed, such as credits not appearing or spend limits being applied incorrectly, but not why, and none of the 11 multi-model incidents gives a cause.

That is not the whole of Anthropic’s disclosure. Its Engineering page (retrieved 2026-10-10) has two postmortems we found, one from 23 April 2026 about three changes that degraded the quality of Claude Code, the Agent SDK and Cowork, and one from 17 September 2025 about three infrastructure bugs that degraded responses. Both are about response quality, not availability, so none of the 50 incidents here has a postmortem that we found. OpenAI, by contrast, links short write-ups from some of its incidents, a difference we compared in our review of OpenAI’s write-ups. We checked Anthropic’s status feed and Engineering index only.

What this means if you rely on Claude

• Allow for hours when the title says several models, but do not assume it. 7 of 11 such incidents stayed open for over two hours, while the other four closed in 26 to 98 minutes. All incidents naming one to three models cleared within 95 minutes on the page.
• If your work depends on one model, check that model’s recent incidents. Incidents naming specific models were common (18 of 50) but short.
• Read the title, then the impact label. 22 of 50 were major or critical. Product incidents were labelled that way about as often as one-to-three-model ones (7 of 17 and 8 of 18), against 7 of 11 for multi-model ones.
• Do not expect a cause. In 47 of 50 incidents the status page does not attribute the problem to anyone, so plan around the duration and not the explanation. Our study of status-page stages shows how much of an incident’s open time can come after Anthropic says a fix is in.

What we could not verify

• We cannot tell how many users were affected by any incident, so the dataset measures what the page displays, not user impact. Twelve resolved updates give their own impact windows, and several differ from the page time.
• The five groups are our reading of incident titles, written and sometimes edited by Anthropic’s staff, and the feed shows final titles. A differently worded title could move an incident between groups. The 17 product incidents are a mixed group.
• Anthropic’s feed holds only its 50 most recent incidents, from 27 July, so we cannot say how this period compares with earlier ones, and the monthly counts cover unequal spans (July only from the 27th, October only to the 7th).
• Durations run from the incident’s creation to its first resolved update and include any monitoring stage. The 5,961-minute Windows incident is 58% of all minutes, which is why the mean is 210 and the median 63. One incident was still in monitoring when we read the feed.
• The weekday and hour patterns describe when incidents were logged, not necessarily when problems began, and the hour window was chosen after looking.
• We do not know what caused any of the multi-model incidents, so the duration pattern is an observation, not an explanation. Our flag for third-party attribution matches three specific phrases, so it counts only those.
• We checked Anthropic’s status feed and Engineering index only, so postmortems published elsewhere would not appear here.

How we built the dataset, and how you can check it

On 10 October 2026 we saved Anthropic’s incident feed, which is Atlassian Statuspage’s JSON, with every incident’s updates and component statuses. A script (dataset_anthropic.py) took each incident’s start and first resolved time, impact label and affected components, flagged incidents whose updates attribute the problem to a third party, assigned the five groups from the title using the rules in the script, and printed every figure in this article. The per-incident data, with start, end, minutes, impact, components, the flag and our group, is in this CSV file (50 rows, retrieved 2026-10-10), so you can recompute the figures. The feed rolls forward, so the CSV is the record of what we measured. For OpenAI’s equivalent, see our dataset of 111 OpenAI incidents, and for the two providers’ counts side by side, our comparison and its caveats.

Data dictionary and how to cite

CSV column	Meaning
id, title	Anthropic’s incident identifier and the incident’s final title
start_utc, end_utc, minutes	Creation time and first resolved time, UTC, and the minutes between them (blank if unresolved)
impact	Anthropic’s own incident label: none, minor, major or critical
components	Components an update set to a non-operational status, separated by semicolons
stated_cause	True if an update attributes the problem to a third party (three phrases in the script)
category	Our grouping from the title: one model, two or three models, several models, claude product, or billing

Cite as: Our Nation Online, “Anthropic status incidents, 27 July to 7 October 2026” (CSV of 50 incidents, retrieved 2026-10-10), https://ournationonline.com/wp-content/uploads/2026/10/anthropic-status-incidents-2026-07-27-to-2026-10-07-v2.csv. We have not applied a licence to the data, so please link back to this page when you use it.

Frequently asked questions

If something is failing right now, our step-by-step guide shows what to check first: ChatGPT or Claude not working? A five-step check.