An hour, and some messages that are not there
Anthropic opened an incident on its status page at 14:21 UTC on Monday, reporting elevated error rates across claude.ai, Claude Code, Claude Cowork and the Claude API. Errors had begun around 14:00 and hit chats, sign-ins and Claude Code sessions.
The company applied a mitigation at 14:41 and error rates dropped, but a second failure followed at about 15:00 that blocked sign-ins — including single sign-on and Sign in with Apple — new chats, voice conversations, Claude Code and Cowork sessions, purchases and file uploads. Services were declared operational again at 15:30.
The line that matters is the one Anthropic added afterwards: “Some messages sent between 14:00 and 14:59 UTC may not have been saved.”
Losing the message is not the same as losing the minute
Outages at this scale are routine and mostly uninteresting. An hour of elevated errors at a frontier lab is a Monday.
Losing user content is different in kind. A failed request returns an error and the user retries; an accepted message that was never persisted looks to the user like it worked, and the loss is only discovered later, if at all. Anthropic has not said how many messages are affected, and its wording — “may not have been saved” — suggests it does not have a precise count.
For anyone using Claude Code or Cowork as a working record rather than a chat, that is the operationally relevant fact, not the 89 minutes of downtime.
The timing
Anthropic is preparing a listing. Its IPO prospectus, reported last week, discloses $4.6 billion of revenue against $42 billion of losses and $518 billion of outstanding cloud infrastructure commitments, and warns investors that its own technology could pose existential risks.
Reliability disclosure is a live question for any company in that position. A one-hour incident changes nothing about the filing. A pattern of them, disclosed after the fact and without counts, is the kind of thing enterprise buyers start asking about during a diligence period.
What to watch
Anthropic has not published a post-incident review for this outage. Whether it does — and whether it quantifies the unsaved messages rather than leaving the qualifier in place — is a small, concrete test of what its disclosure practice looks like as a public company.
