Anthropic's Claude Opus 4.7 outage on May 12, with two downstream providers catching it in real time

Anthropic reported elevated errors on Claude Opus 4.7 at 23:38 UTC on May 12, 2026. Four components were affected: claude.ai, the Claude API at api.anthropic.com, Claude Code, and Claude Cowork. By 23:58 UTC the incident was resolved. Total duration: 19 minutes.

The vendor's own timeline was thin. Three updates, no root cause disclosed.

The issue has been identified and a fix is being implemented.

That came at 23:38 UTC, simultaneous with the incident opening. Twelve minutes later:

A fix has been implemented and we are monitoring the results.

Then resolution at 23:58 UTC with no further detail. Anthropic didn't say which layer broke, what the error rate looked like, or which requests were affected. Users on Claude Code and the API would have seen failures; how many and for how long within that 19-minute window isn't in the public record.

The cascade

Two downstream providers picked this up in real time and named Anthropic explicitly in their own status updates.

Relevance AI posted at 23:59 UTC, one minute after Anthropic resolved:

Anthropic recovered Anthropic went down

Looks like a status-bot dump, but the content is clear enough. Their incident opened and closed at the same timestamp (23:59 UTC), which suggests an automated dependency check that fired on resolution rather than on the initial failure. Worth noting: Anthropic's incident had already closed by the time that update appeared.

bolt.new was similar, timestamped 23:58 UTC:

[Dependency Check] Anthropic Status: Identified Upstream Provider has reported an outage Affected components Anthropic (Degraded performance) Status: Resolved This incident has been resolved.

Same pattern. Automated dependency monitor, probably polling Anthropic's status page, caught the tail end of the incident as it cleared. bolt.new's incident also opened and closed at the same timestamp.

Neither downstream provider had a user-visible outage of their own that extended beyond Anthropic's window, at least not one they reported. Both treated it as a dependency notification rather than an incident requiring investigation.

Some context on frequency

Anthropic has had 85 incidents in our database over the past 12 months, 34 of which we classified at the same severity level as this one. Median incident duration across that set is 50 minutes. At 19 minutes, this one closed faster than most. There was also a separate Anthropic incident earlier the same day, starting at 19:36 UTC, so May 12 had two back-to-back events.

What neither Anthropic nor the downstream providers explained is why Opus 4.7 specifically was affected while other model variants apparently weren't. The component list covers all the user-facing surfaces (web, API, CLI, Cowork) but doesn't distinguish by model, so it's unclear whether other models were degraded at a lower rate or were entirely clean. The vendor didn't say.