AI Pulse · Frontier Labs Watch

Anthropic is investigating problems across multiple Claude models today, including Opus 5, Sonnet 5, and Haiku 4.5. This comes on the heels of a separate outage on August 16, where Downdetector logged more than 2,000 reports from users. That’s three degraded-performance incidents in under a week, and if you’re running production work through Claude Code, that pattern matters more than any single outage does. Details are in the incident status page.

Here’s the tension worth sitting with. At the same time Anthropic is dealing with repeated reliability wobbles, the company is actively courting enterprise trust with watermarking features, EU compliance work, and sovereign compute deals. Those are the kinds of things you pitch to a CFO or a compliance officer to convince them your platform is stable enough to build on. A third incident in a week doesn’t line up with that pitch, and it’s fair to ask whether the infrastructure is keeping pace with the sales conversation.

For an owner or operator, the practical question isn’t whether Claude is a good model. It clearly is, and plenty of teams rely on it daily. The question is whether you have a plan for when it goes sideways mid-task. If you’ve got Claude wired into a workflow that customers or employees depend on, you need a fallback path, a way to detect degraded output before it reaches someone downstream, and a clear view of which model version is running where. That kind of monitoring and failover logic is exactly the kind of thing we build into an AI command centre, so a hiccup at the model layer doesn’t turn into a hiccup in your business.

Nothing here is catastrophic yet. But three incidents in a week is a trend, not a fluke, and it’s worth watching closely before you add more weight to that foundation.