Claude went down twice in 14 hours — and Anthropic's uptime numbers tell the full story

Two separate Claude outages hit within 14 hours of each other on June 22 and 23, and the numbers sitting on Anthropic’s own status page make the situation harder to ignore. Across claude.ai, the Claude API, and Claude Code, 90-day uptime figures are currently 99.12%, 99.41%, and 99.28% respectively. That sounds fine until you compare it to the 99.9% threshold that most enterprise software contracts require. At 99.9%, you’re allowed roughly two hours of downtime per service per quarter. Anthropic’s current figures translate to somewhere between 19 and 23 hours per service. That gap matters if your team is building production workflows on top of Claude.

What actually happened

The first incident began just after midnight UTC on June 22. Anthropic’s engineering team opened an investigation at 00:37 UTC into elevated error rates hitting five models simultaneously: Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5. Recovery didn’t happen all at once. The team brought models back one by one, with monitoring confirming a fix by 02:06 UTC. A roughly 90-minute window of disruption, resolved before most US-based users started their day.

Then a second wave emerged the same afternoon. Elevated error rates across the same model set were logged from 08:11 UTC, with a fix confirmed at 14:44 UTC on June 22.

Less than 14 hours after that, the June 23 incident began. Reports started coming in around 10:02 am Eastern Time, and Downdetector peaked at over 8,000 reports in the United States alone. Anthropic’s status page logged the investigation opening at 14:19 UTC and marked the issue as identified by 14:25 UTC, six minutes later. Error rates broadly returned to normal by 16:44 UTC. Affected services spanned the full platform: Claude.ai, the Claude API, Claude Console, Claude Code, and Claude Cowork.

Anthropic has not published a root-cause analysis for either incident.

The demand problem Anthropic acknowledged publicly

When asked about the pattern of disruptions, Anthropic told Fortune: “Demand for Claude has grown at an unprecedented rate, and our infrastructure has been stretched to meet it, particularly at peak hours.” The company pointed to compute capacity deals with Amazon and Google as sources of relief, though most of that capacity isn’t expected to be operational until late 2026 or into 2027.

The scale of the mismatch is striking. Anthropic’s annualised revenue run rate moved from roughly $9 billion at the end of 2025 to $47 billion by May 2026. The company has separately disclosed that demand is running approximately 80 times higher than it planned for at the start of 2026. Infrastructure cannot scale that fast, regardless of how many supply contracts you sign.

These two incidents on June 22 and 23 are not isolated. TechTimes documented what was at the time the tenth significant disruption since June 5, averaging roughly one per day. The June 22 overnight outage was the third disruption that month alone, following a major multi-hour event on June 2 and a networking degradation incident on June 5.

What the error types tell you

Anthropic’s API documentation distinguishes between two error types developers encounter during outages. An HTTP 500 error means something has failed inside Anthropic’s infrastructure and requires a server-side fix. An HTTP 529 error means the system is technically healthy but is rejecting requests because it has hit capacity limits. Crucially, 529 errors do not count against your usage quota and resolve when capacity frees up. During peak-demand outages, knowing which error you’re seeing tells you whether to wait it out or escalate.

What this means if you’re building on Claude

If your team uses Claude primarily through the chat interface for research or drafting, a 90-minute outage is annoying. If you have production pipelines calling the Claude API, or CI/CD workflows integrated with Claude Code, the calculus is different. Development cycles stall, automated content pipelines miss scheduled runs, and any customer-facing feature that depends on Claude goes silent.

The June 2 incident is worth understanding as a specific warning. That outage was traced to a bug in Claude Code’s sub-agent architecture, where sub-agents designed to parallelise complex coding tasks entered an infinite loop and multiplied exponentially. Anthropic issued an automated quota reset for affected Pro and Max accounts after the fact. But if your pipelines are deeply integrated with Claude Code’s agentic features, you are inheriting whatever failure modes exist in that architecture, and Anthropic has not yet published detailed post-mortems explaining what changed to prevent recurrence.

Practical steps worth taking now, if you haven’t already: implement multi-model failover routing so a single provider outage doesn’t halt your entire pipeline, treat AWS Bedrock’s Claude access as a separate failure domain from Anthropic’s direct API, and add circuit breakers to any workflow that calls Claude Code as part of an automated process.

The broader timing

All of this is happening roughly three weeks after Anthropic confidentially filed for an IPO that could value the company above $1 trillion. Operational reliability is not a minor footnote for a company at that stage. Investors and enterprise procurement teams making multi-year commitments will be looking at the same uptime figures that are publicly visible on status.claude.com, and those numbers currently sit below the threshold that enterprise SLAs typically require.

Anthropic’s compute partnerships with Amazon, Google, Microsoft, and others represent genuine capacity on the way. But the gap between contracted capacity and operational capacity is measured in months, and the current demand curve is not waiting.

For now, the practical reality is that Claude is a powerful platform running under significant strain. Plan your integrations accordingly.