Claude Outage History: 382 Incidents So Far in 2026, Up From 288 a Year Earlier

Claude outage history, as of November 2, 2026: status.claude.com records 934 incidents since March 14, 2023, the latest on October 28. It logged 382 from January 1 to November 1, 2026, up from 288 a year earlier and already above all of 2025. Monthly counts from August to October are 63% below February to July.

We downloaded all 938 entries from the JSON behind Anthropic’s status page, removed 4 scheduled maintenance notices, and joined the records to Anthropic’s own per-component outage calendars. This is Anthropic’s record of itself: useful for dates, scope and trend, but not an independent measurement.

The latest entries in the record

The 14 most recent incidents as of November 2. Nothing has been posted since October 28, and the page read “All Systems Operational” when we pulled it. Times are UTC and each title links to Anthropic’s own entry.

That list covers only 21 days, and it would have covered far fewer in early summer. July 2026 had 54 incidents; September had 13 and October 17.

Claude down history, year by year

*2026 runs to November 1. Each incident is counted once at its worst component status; 152 incidents with no component status are in the total only. Outage hours are partial plus major hours from Anthropic’s daily uptime calendars.

The record starts with two “Elevated API Errors” entries on March 14 and 15, 2023, where Anthropic said a majority of API traffic was hit for one hour and then for 30 minutes. The page logged 30 incidents that year, 188 in 2024 and 334 in 2025. Same-period 2026 is up 33% on 2025, while new components arrived along the way: Claude Code in May 2025, Claude for Government and Claude Cowork in 2026. More components and more models give the page more to report, so these counts mix reliability with coverage.

The two panels share a time axis but have separate scales. Hover or tap for values. March 2026 is the peak in outage hours: claude.ai 39.0 h and the API 39.4 h.

How often is Claude down?

Often, if any logged issue counts, and much less often if you only count outages. In the 305 days to November 1, 2026, the status page recorded an incident on 187 days, or 61%. An incident reached partial outage or worse on 92 days (30%) and major outage on 18 days (6%). Anthropic’s calendar shows claude.ai with some partial or major outage time on 89 days and the API on 77.

Uptime is computed from Anthropic’s daily calendars using Atlassian Statuspage’s published rule: major outage minutes count in full, partial outage minutes at 30%, and degraded performance not at all.

The drop since August is concentrated in the mildest entries. Degraded-only incidents averaged 28 a month from February to July and 9 from August to October, while incidents reaching partial outage or worse went from 16 to 9. Incident titles also started bundling more models: entries naming a model listed 1.2 models on average from February to July and 2.3 from August to October. So some of the fall could be consolidation or a higher bar for posting. Outage hours fell as well, with claude.ai averaging 19.9 hours a month in the earlier period and 6.3 in the later one, and the API 17.1 and 5.3. The status data cannot say how much of that is better service and how much is reporting.

By product

An incident can list several components. Claude Cowork’s status component starts April 1, 2026 and Claude for Government’s February 17, 2026. 25 incidents in 2026 list no component.

The Claude Code outage record

Claude Code has its own status component, and the first incident listing it is dated May 22, 2025. 103 incidents listed it in 2025 and 285 did from January 1 to November 1, 2026. 103 of the 2026 entries reached partial outage or worse on Claude Code and 16 reached major outage. Anthropic’s calendar gives 115.4 hours of partial or major outage in that period, against 33.9 in the same months of 2025. Its monthly reported uptime ranged from 98.57% in March to 99.85% in October.

Most Claude Code outages are not specific to Claude Code. 250 of the 285 entries in 2026 also list the API, and only 11 list Claude Code alone. 15 mention login, sign-in or authentication. The ones that are specific to the tool are often client releases: version 2.1.59 caused file write contention for Windows users on February 26, and version 2.1.120 crashed when resuming a session on April 25, with auto-update moving clients back to 2.1.119.

Anthropic’s April 23, 2026 engineering post covers a different kind of problem. It traced reports of worse Claude Code answers to three changes between March 4 and April 20 (a lower default reasoning effort, a bug that cleared session thinking, and a system prompt instruction to reduce verbosity) and says the API was not impacted. We searched March and April 2026 incident titles and updates for those issues and found no matching entry, so the status record, and every count on this page, misses them.

Which Claude models appear most

534 of the 934 incidents (57%) name a specific model. The most named are Sonnet 3.5 (92), Sonnet 4 (56), Opus 4.6 (45) and Haiku 4.5 (39). By family, Sonnet leads across all years with 275 mentions against 239 for Opus and 79 for Haiku, but the mix flipped: Sonnet went from 145 mentions in 2025 to 75 in 2026, and Opus from 64 to 154. A count is not a failure rate, since we have no request volumes by model and long-lived models had more time to collect entries.

A model is counted when its title names it. When the title is generic, or says “and other models”, we count models named in the updates. Launch and retirement dates come from Anthropic’s release notes and model deprecations page; Claude 3 launch dates are not on those pages, so those tracks start at the first incident.

New models collect incidents fast

For the 17 models launched since February 2025 with at least 30 days of history after their first month, incidents naming the model ran at 8.2 per 30 days in the first 30 days and 2.5 per 30 days after that. 14 of the 17 had a higher rate in the first month. Opus 4.8 is the extreme, with 23 incidents in its first 30 days. Launch demand and closer reporting of a new model would both produce this pattern, and the data cannot separate them. The newest models have started more quietly: Opus 5.5, Sonnet 5.5 and Haiku 5.5 logged 1, 4 and 2 in the time they have had so far, during months when the page logged fewer incidents of every kind.

Rates stop at the model’s retirement date or November 1, 2026. Fable 5 and Mythos 5 were suspended for 18 d 18 h 36 min of their first month, and that suspension counts as one incident for each.

A model in the title does not mean a narrow listing. 194 of the 215 model-titled incidents in 2026 list three or more components, and 36 of the 184 that name exactly one model marked the Claude API as partial or major outage. Anthropic has no per-model component, so a problem with one model can register against the entire API.

When Claude outages start: hour of week

Incidents that began in each hour: fewer more. Top row is the hour of day.

Across 684 incidents from March 2025 to October 2026, 90% began Monday to Friday, against 71% if incidents were spread evenly, and 201 of 226 partial or major outages did too. The busiest weekday is Tuesday (147) and the quietest day is Sunday (28). The busiest hour is 9:00 to 9:59 Eastern (60 incidents); on a UTC clock it is 14:00 to 14:59 (59). 40% began between 9 a.m. and 3 p.m. Eastern. The busiest cell in the grid is Monday at 9:00 Eastern, with 16.

There is a second, smaller cluster that a US-only view misses: 136 incidents (20%) began between 2 a.m. and 6 a.m. Eastern, which is 06:00 to 10:00 UTC in summer. We have no traffic or deployment data, so we cannot test whether either cluster follows demand, engineering changes or something else. Start times here are also the first post, not the first failure.

How long Claude outages last

The longest and widest Claude outages, and the causes Anthropic stated

Each row is an entry notable for length or reach. “Components at worst” counts how many status components carried the incident’s worst status. The last column summarizes what Anthropic wrote; where it named no cause, we say so.

The longest entry in the record was not a technical failure. On June 12, 2026 a US export control directive required Anthropic to restrict foreign nationals’ access to Claude Fable 5 and Mythos 5. Anthropic says it had no reliable way to verify nationality in real time, so it suspended both models for all users. Access to other models was not affected, and the entry closed on July 1 after the controls were lifted. The status page marked it critical in impact but only degraded at component level, so it adds nothing to the outage hours above. Without it, the longest entry is 8 d 20 h 24 min, “Model output quality” in September 2025.

The widest outages by that measure are April 24 and April 30, 2026, when all six components were marked as major outage for 39 min and 31 min. 4 incidents marked five or more components as major outage, all in 2026, after Claude for Government and Claude Cowork joined the four core components. Neither April entry states a cause.

Outside events show up in the record too. Anthropic attributed June 12, 2025 to “an ongoing outage impacting GCP”. On November 18, 2025 it blamed an unnamed “third-party vendor”, and on December 5, 2025 an unnamed “upstream provider”. Those two entries fall inside or overlap Cloudflare’s November 18 outage (11:20 to 17:06 UTC) and its December 5 outage (08:47 to 09:12 UTC). That is a match on timing only, since Anthropic does not name Cloudflare.

Why Claude goes down, as far as the record says

Mostly the record does not say. We read every update matching a search for cause-related words and coded those that name a trigger: 37 of 934 incidents (4%), and 9 of the 58 that reached major outage. 11 entries carry a post-incident note in the status record’s post-mortem field, nearly all from December 2025 to March 2026. Within that sample, 19 of 37 (51%) name a change Anthropic made, and 9 name an outside provider. The sample is whatever Anthropic chose to explain, so it does not show the causes of the other 96%.

Is Claude getting more reliable?

Year on year, the status record mostly says no. From January 1 to November 1, incidents rose from 288 in 2025 to 382 in 2026 and partial-or-worse incidents from 102 to 128. claude.ai outage hours rose from 112.5 to 142.3 and API hours from 82.5 to 122.0. Two measures moved the other way: incidents reaching major outage fell from 25 to 21, and median duration from 1 h 14 min to 1 h 2 min.

Over the last three months, it says yes. Incidents, days with an incident and outage hours all fell from August, and claude.ai’s reported uptime reached 99.85% in October, its best month since September 2025. Three caveats apply. Three months is short: claude.ai’s reported uptime also rose from 98.07% in July 2025 to 99.80% in August 2025, then fell back to 99.22% by November. Titles started bundling more models over the same months. And there is no denominator: Anthropic does not publish request volumes on the status page, so we cannot tell whether errors per request rose or fell.

On that reported-uptime measure, since January 2025 claude.ai reached 99.9% reported uptime in 1 of 22 months and 99.5% in 11. The API reached 99.9% in 0 and 99.5% in 15. Our 99.9% uptime explainer sets out what each level allows; 99.9% is about 44 minutes of downtime in an average month.

Claude and OpenAI on the same definitions

Because our OpenAI outage history uses the same kind of data, we applied one set of definitions to both raw datasets for January 1 to October 29, the last day of the OpenAI pull. Start is the first post or component impact, end is the first “resolved” update, and worst status is the highest component status in the incident.

On those figures, Claude’s page logged more incidents in 2026 than OpenAI’s (382 to 300), after the same number in the 2025 window (284 each). Its incidents more often reached partial outage or worse, carried more outage hours, and were shorter, with a median of 1 h 2 min against 1 h 44 min.

Status pages are self-reported, and these numbers are not directly comparable. Each company decides what counts as an incident, when to post it and which label to use. The structures differ too: OpenAI’s page has 38 components, including 14 for APIs, and marks problems against narrower parts, while Anthropic has 6 and no per-model component, so a single-model problem can mark the whole Claude API. 36 entries in 2026 did exactly that. Merging OpenAI’s ChatGPT components by group, as in the table, raises its 2026 figure from 16.5 hours for the single Conversations component to 59.8, which shows how much the structure matters, but it cannot make the two pages equivalent. The pages also serve very different audiences, as our ChatGPT market share report shows, and neither publishes request volumes. Use the table to see how each vendor reports, not to rank them.

What this means for MSPs

Quote hours and surfaces, not incident counts. A client who reads “382 incidents” will assume more risk than a client who reads 142 hours of partial or major outage on claude.ai over ten months, mostly partial. Neither is the client’s experience, but the second can be set against the client’s own cost of downtime. Keep degraded performance in view as well: it does not count toward reported uptime at all.

Plan the app and the API separately. In 2026, 65 incidents listed claude.ai without the API, and 17 the API without claude.ai; 250 listed both. Anthropic’s note on March 11, 2026 is a typical split: its primary application database degraded, sign-ins for Claude Code and the Console failed, and API traffic was unaffected. A client on Claude seats and a client running an integration need different fallbacks.

Treat sign-in as its own dependency. 20 incidents in 2026 mention login, sign-in or authentication, and 15 of them reached partial outage or worse. During the September 29, 2026 incident Anthropic told signed-in users, “please don’t sign out”. Scripts and agents using API keys may keep working when interactive sign-in does not, so check which path each client workflow uses.

Retries cover blips, not incidents. Anthropic’s API documentation lists 500 and 529 errors, advises exponential backoff for 500s, and says its SDK retries transient failures twice by default. With a median incident of 1 h 2 min, that will not carry a workflow through. Add a circuit breaker that fails over after a set error rate, and handle errors that arrive mid-stream, which the documentation says can occur after a 200 response.

Build failover in layers, and test it. A second Claude model may help when an incident names one model. A different cloud route may help in some cases: the September 2025 bugs hit Anthropic’s own API and Amazon Bedrock at very different rates. It is not free, though. On January 23, 2026, Anthropic’s stated mitigation for invalid thinking signatures was to “not mix Claude API traffic between Bedrock, Vertex, and api.anthropic.com”, so moving a live conversation between platforms can itself produce errors. Provider-level events and the 4 widest incidents need a second vendor. Prompts, tools and output checks all need testing against that vendor before an outage.

Monitor the feed and run your own check. Subscribe to status.claude.com/history.rss or /history.atom, or to the page’s email, SMS or webhook updates, and open a PSA ticket automatically. Do not make it the only signal. In the 24 incidents where Anthropic wrote “Impact occurred from” a time on the same or previous day, the first post came a median of 19 minutes after that time, and more than 15 minutes after it in 14. On August 14, 2026 the status page itself was hard to reach because of an invalid certificate. A small synthetic API call from your own monitoring catches both cases.

Do not promise an SLA the vendor does not offer. Anthropic’s documentation calls the standard tier “best-effort availability”. Priority Tier “targets 99.5 percent uptime”, but commitments are no longer sold and that page is marked legacy. The public Commercial Terms provide the services “AS IS” and “AS AVAILABLE” and contain no uptime figure. If a client needs a number, get it in a contract, and match your own service levels to what you actually hold.

Control Claude Code versions on managed machines. The February 26 and April 25, 2026 entries were both client releases, and the April fix included pinning with claude install 2.1.119. For fleets, test updates in a ring first. The May 6, 2026 entry is worth a note in network documentation too: an infrastructure change altered Anthropic’s outbound IP addresses, which Anthropic said could break Claude Code remote sessions for organizations that restrict GitHub access by IP.

Is Claude down right now?

This page is a dated history, not a live monitor. For the current state, use status.claude.com; the older address, status.anthropic.com, redirects there. It lists claude.ai, the Claude Console, the Claude API, Claude Code, Claude Cowork and Claude for Government separately. If the page shows everything operational and Claude still fails for you, compare another model, the web app against the desktop app, and whether you can sign in, since many entries affect only one of those.

Search every Claude incident

All 934 incidents. Search titles and model names, filter by year, worst status or component, and sort by date or length.

Frequently asked questions

How often does Claude go down?

Anthropic’s status page logged 382 incidents from January 1 to November 1, 2026, on 187 of 305 days. 217 of them were degraded performance only. 92 days had an incident reaching partial outage or worse, and claude.ai recorded partial or major outage time on 89 days. From August to October the page logged 18 incidents a month, down from 50 a month in February to July.

Has Claude Code had outages?

Yes. 285 incidents listed Claude Code from January 1 to November 1, 2026, against 103 in all of 2025. 16 reached major outage on that component and 103 reached partial outage or worse, for 115.4 hours in Anthropic’s uptime calendar. 250 of the 285 also listed the API; 11 listed Claude Code alone.

What was the longest Claude outage?

The longest status-page entry ran 18 d 18 h 36 min from June 13, 2026: Anthropic suspended Claude Fable 5 and Mythos 5 for all users after a US export control directive, and other models were not affected. The longest entry with a technical cause is the 8 d 20 h 24 min September 2025 ‘Model output quality’ incident, a degraded-quality issue. The widest outages, April 24 and April 30, 2026, marked all six components as major outage for 39 min and 31 min.

Is Claude getting more reliable?

The data points both ways. Same-period incidents rose from 288 in 2025 to 382 in 2026, and claude.ai partial or major outage hours rose from 112.5 to 142.3. Since August the trend reversed: claude.ai averaged 6.3 outage hours a month in August to October against 19.9 in February to July. Three months is a short sample, and these are self-reported figures with no traffic denominator.

Does Claude go down more than ChatGPT?

On the same definitions from January 1 to October 29, 2026, Claude’s status page logged 382 incidents and OpenAI’s logged 300, and Claude’s reached partial outage or worse 128 times against 39. Claude’s median incident was shorter, 1 h 2 min against 1 h 44 min. Status pages are self-reported and the companies define and split incidents differently, so these counts are not a direct reliability comparison.

Does Anthropic publish an uptime SLA?

Its documentation says the standard API tier has best-effort availability and that Priority Tier targets 99.5 percent uptime, but Priority Tier commitments are no longer sold. The public Commercial Terms provide the services as is and as available and contain no uptime figure. Enterprise contracts may differ; we can only report the public documents.

Methodology

References

1. Anthropic, Claude Status, incident history. Accessed November 2026.
2. Claude Status, history JSON (paged, plus one JSON record per incident). Accessed November 2026.
3. Claude Status, incident history RSS feed. Accessed November 2026.
4. Claude Status, uptime calendar for the Claude API (one per component). Accessed November 2026.
5. Atlassian, Display historical uptime of components. Accessed November 2026.
6. Anthropic Engineering, A postmortem of three recent issues. Accessed November 2026.
7. Anthropic Engineering, An update on recent Claude Code quality reports. Accessed November 2026.
8. Anthropic, Statement on the directive to suspend Fable 5 access. Accessed November 2026.
9. Anthropic, Redeploying Claude Fable 5. Accessed November 2026.
10. Claude Platform Docs, release notes. Accessed November 2026.
11. Claude Platform Docs, model deprecations. Accessed November 2026.
12. Claude Platform Docs, errors. Accessed November 2026.
13. Claude Platform Docs, service tiers. Accessed November 2026.
14. Anthropic, Commercial Terms of Service. Accessed November 2026.
15. OpenAI Status, incident history, retrieved October 30, 2026 for our OpenAI report. Accessed October 2026.
16. Cloudflare, Cloudflare outage on November 18, 2025. Accessed November 2026.
17. Cloudflare, Cloudflare outage on December 5, 2025. Accessed November 2026.