Blog

Anthropic’s own status page says a Claude Cowork bug that lasted from 10 to 14 September 2026, about four days, was fixed by a Windows update from Microsoft. We read that update’s official release notes. Microsoft’s page lists three fixes: Remote Desktop Services instability, a Hyper-V folder-sharing bug, and a multichannel USB audio issue. It does not mention Claude, Cowork, or Anthropic anywhere. That does not mean Anthropic is wrong, since a Remote Desktop Services fix could plausibly affect Cowork’s own remote session handling, but it is Anthropic’s own attribution, not something Microsoft’s documentation confirms.

Key takeaways

1. Anthropic’s status page logged 13 incidents in September 2026 (1-29). We read every one, in full, on status.claude.com.
2. One of them, degraded Claude Cowork on Windows, stayed open for about 4 days: about 34 times longer than the next-longest single incident that month, and about 7 times the combined length of all 12 other incidents put together.
3. Anthropic says a specific Microsoft update, KB5129195, fixed it. We read that update’s own release notes and found no mention of Claude, Cowork, or Anthropic.
4. Two other incidents show a small gap between the headline timestamp and the impact window described in the incident text, a pattern worth knowing about before you trust either number exactly.

Last reviewed 30 September 2026. We read Anthropic’s full incident history for September 2026 and Microsoft’s KB5129195 release notes in a browser on that date. We did not experience or reproduce the Windows bug ourselves.

What did Anthropic log in September 2026?

We read all 13 incidents listed on Anthropic’s status history for 1 to 29 September 2026, retrieved 30 September 2026. Twelve were short, most under 3 hours, covering elevated errors on specific models (Claude Mythos 5.1, Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5), Claude for Microsoft Office 365, Google Play subscription billing, and credit purchase delays. The thirteenth, the Windows Cowork bug, was different in kind, not just length.

Add up the twelve short incidents and they total roughly 14 hours 45 minutes for the whole month. The Windows Cowork incident alone ran about 99 hours: about 7 times that combined total, and about 34 times the length of the month’s next-longest single incident (the 3 September multi-model error at just under 3 hours).

The bug, and the fix Anthropic pointed to

Anthropic’s own description of the incident reads: “This issue has been resolved. Microsoft has released a September 14 Windows update that fixes it,” with a direct link to Microsoft’s KB5129195 page. We followed that link and read the release notes ourselves (KB5129195, Microsoft Support, retrieved 2026-09-30). Microsoft’s own title labels it an “Out-of-band” update, meaning one released outside the normal monthly schedule. Its Remote Desktop Services fix is described as addressing an issue that appeared “after installing the September 2026 Windows security update (KB5122880)”, so KB5129195 is itself a follow-up patch. Its listed fixes are:

• Remote Desktop Services instability that could cause RDP connections and sign-ins to fail, or affected servers to become unresponsive.
• A Hyper-V bug where folders shared from a Windows host to a Linux virtual machine using Plan9 would not appear or could not be accessed.
• A problem with some USB Audio Class 1.0 devices failing in multichannel or 8-channel and 3D audio modes.

Nowhere on Microsoft’s page is Claude, Cowork, or Anthropic mentioned. The Remote Desktop Services fix is the most plausible link, since Claude Cowork’s own remote and session-based features could plausibly depend on the same Windows services RDS uses, but that link is our own inference, not something either company states directly. Anthropic did not publish its own explanation of the underlying cause beyond attributing the fix to Microsoft’s update.

A small but real pattern: which timestamp do you trust?

Two of the thirteen incidents show a gap between the bold headline timestamp Anthropic gives the incident and the plain-text description of when the impact actually occurred. The 11 September incident is headlined “14:19 – 14:37 UTC” but its own description says “Impact occurred from 13:57 to 14:20 UTC”, a start time 22 minutes earlier than the headline. The 2 September incident is headlined “21:17 – 21:44 UTC” but describes impact as “2:05pm PT / 21:05 UTC to 2:19pm PT / 21:19 UTC”, a start about 12 minutes earlier and an end about 25 minutes earlier than the headline. We do not know which figure Anthropic considers authoritative. If you need an exact duration for one specific incident, we would read both the headline and the description text, not just one.

What this means if Claude breaks for you

• Check status.claude.com first. Most incidents in our count cleared in under 3 hours.
• If it is Claude Cowork specifically and you are on Windows, a pending Windows update is worth checking, on the evidence of this one incident. It is not a general rule, since we only have one documented case.
• Read the incident’s own description text, not just the headline time, given the gap we found in two of the thirteen September incidents.
• A short incident on the status page does not rule out a longer personal problem. As with our earlier count of OpenAI’s status incidents, a status page reports aggregate monitoring, and an account-specific issue may not appear as a public incident at all.

What we could not verify

• We did not reproduce the Windows Cowork bug or test the fix ourselves.
• We cannot confirm Anthropic’s attribution of the fix to KB5129195 beyond Anthropic’s own statement, since Microsoft’s release notes do not name Claude.
• We do not know how many Cowork-on-Windows users were actually affected during the four days; Anthropic’s page gives no user count.
• This count covers 1 to 29 September 2026 as read on 30 September. Anthropic’s page may add or amend entries after this was written.

For the same exercise done on a different provider, see our case study of a ChatGPT Work Mode incident marked resolved while users still reported it broken. Both pieces use the same method: read the primary source in full, quote what it actually says, and flag the gap between what a status page claims and what happened.

If you use Claude alongside other assistants, see our comparison of ChatGPT, Claude, Gemini and Perplexity.

Why this kind of cross-check matters

A status page is written by the company whose service broke. That does not make it dishonest, but it does mean the convenient explanation, in this case, blaming an outside vendor’s update rather than Anthropic’s own code, is worth checking against a second, independent source before repeating it as fact. Here, the independent source, Microsoft’s own release notes, neither confirms nor contradicts Anthropic’s claim; it simply does not address it. That is a different finding from either “confirmed” or “false”, and reporting it as anything stronger than that would overstate what we actually found.

How we researched this

On 30 September 2026 we opened Anthropic’s status history, expanded September in full, and recorded every incident’s title, stated window, and description text. We then opened the linked Microsoft KB page directly and read its listed fixes in full. Where we describe a cause, it is stated on one of these two pages; where we note a gap or an inference, we say so.

Frequently asked questions