At 11:53 UTC on April 28, 2026, Claude Haiku 4.5 began returning errors.
By 23:33 that same day, Anthropic's status page had logged five separate incidents. Haiku 4.5, Sonnet 4.5 and Opus 4.7 each had failures; the login path for Claude.ai was down for 78 minutes from 17:34 to 18:52; and Claude Code's Code Review feature intermittently failed to start sessions.
Taken together, Anthropic services were partly or fully unavailable for nearly two hours across one workday.
What broke this time
Anthropic described the incident as a problem causing elevated Anthropic API errors and trouble reaching Claude.ai, including login paths for Claude Code.
In practical terms, the API was throwing 500 errors, the web app was not letting users in, and Claude Code could not reliably get to the starting line.
Downdetector showed the scale more plainly: by 11:11 a.m. Pacific time on April 28, more than 5,000 users had filed outage reports. This was not a narrow regional hiccup, but a simultaneous global service problem.
| Service | Start | End | Duration |
|---|---|---|---|
| Haiku 4.5 | 11:53 | 12:44 | 51 minutes |
| Sonnet 4.5 | 13:22 | 13:39 | 17 minutes |
| Claude.ai + API | 17:34 | 18:52 | 78 minutes |
| Opus 4.7 | 23:25 | 23:33 | 8 minutes |
| Claude Code Review | Several periods | - | Intermittent |
Five separate failures on the same business day left few clean windows for teams trying to run an uninterrupted workflow.
Not the first time this month
The timing is what makes the episode harder to dismiss. Eight days earlier, on April 20, Claude had suffered a similar broad outage, with 6,500 Downdetector reports mentioning messages such as "Service is temporarily busy" and "Page not found."
Anthropic did not publish a postmortem for either incident.
For a company selling itself as enterprise-grade infrastructure, that is an awkward stance. In enterprise software, an incident that affects thousands of users and spans multiple product lines usually earns customers a formal report with a root-cause analysis.
Why it matters now
Anthropic's 2025-to-2026 story has been built around revenue tripling in three months, a reported 40% share of the enterprise API market and a valuation of $380 billion.
The operational reality behind those numbers is that growth appears to be outrunning infrastructure capacity.
Several signals now stack up: Claude Code had just been slated for removal from the $20 Pro plan in April as pricing was being reworked; on April 24, Anthropic publicly reviewed a recent decline in Claude Code quality and acknowledged engineering mistakes; reliability issues are now arriving on top of those changes.
Put simply, the company is trying to reprice products, iterate models and expand enterprise usage at the same time. Each of those is difficult alone. Together, they push incident risk higher.
What customers will watch
For companies adopting Claude, the decisive question is not only benchmark performance. SLA terms, availability and the speed of incident response are the hard clauses that make it into procurement contracts.
The next few weeks are worth watching for three things: whether Anthropic publishes a postmortem; whether the problem was limited to control-plane systems such as login and API gateways or reached the compute layer; and whether inference infrastructure is keeping pace under the AWS Trainium training partnership.
If those questions go unanswered, the next bargaining chip for large customers such as NEC, CBA and Canva may not be price.
Sources: Claude Status official records (status.claude.com), Anthropic's Claude suffers alarming outage for thousands (Rolling Out), CocoLoop, Claude AI Services Suffer Repeat Outages Impacting Global Users (infonasional.com), Claude AI was down - what we know so far about the outage (Tom's Guide)