Did Claude Sonnet just get dumber?
I've been noticing this for the past two weeks. Clearly wrong answers. Forgetting decisions we'd made ten minutes earlier ("Ah, yes, you're right, we already made that decision"). Building complex documents without being asked to ("sorry I misunderstood what you meant and created the wrong thing"), then apologising for misunderstanding. Not following the structure of a planned task.
It's most obvious on Claude.ai and Claude Desktop. I've moved to using Claude via terminal for about 80% of my work now, and the difference is evident. The terminal version doesn't show any of this. The performance degradation feels like we're back to where we were eight months ago.
So I started looking into whether others were seeing the same thing. They are.
What's actually been happening
March 2026 has been a rough month for Anthropic's infrastructure. Their own status page logged major outages on March 2–3, elevated error rates on March 11 and 12, clustered incidents across March 16–18, authentication failures on March 19, hanging responses on March 20, and Opus/Sonnet errors again on March 21. Roughly one significant incident every two to three days.
On March 26, Anthropic officially confirmed they were tightening 5-hour session limits during peak weekday hours to manage surging demand, following a significant wave of new users after the OpenAI Pentagon contract controversy in late February. Around 7% of users were told to expect limits they hadn't hit before.
A developer investigation published this week pointed at a broken prompt caching system as a likely contributing factor. Tool schema bytes were changing mid-session, invalidating cached conversation history and forcing full token recalculation on every turn. The longer the session, the worse the effect.
Whether all of this is purely infrastructure, or whether some of it reflects deliberate choices about capacity and cost, is an open question. Anthropic says they never intentionally degrade model quality. The timeline is at least worth examining.
Why the terminal behaves differently
The API, which is what the terminal uses, runs on different infrastructure and isn't subject to the same interface-level routing and caching behaviour. It's not that one model is smarter than another. The path from your message to the model differs, and right now some of those paths are less stable than others.
What to do about it
Keep sessions shorter. Restart with a deliberate recap when the reasoning starts to drift. If you need consistent output, the API is more stable than the consumer interfaces right now.
I'll be watching whether this settles once the infrastructure catches up to demand. If it's still happening in a few weeks, that's a different conversation.
I still love you, Claude. But don't make me regret it.