Is Claude Getting Worse? What I Found After Digging Around

Zivdev — Wed, 15 Apr 2026 10:15:04 +0000

There's been a lot of noise this week about Claude being nerfed.

Claims started surfacing on X: median thinking dropped from ~2,200 to ~600 chars, API requests up 80x, and Claude Opus 4.6 reportedly fell from 83.3% to 68.3% on a hallucination benchmark.

Anthropic pushed back. Their response was that it was an intentional token-saving change, made transparently in the changelog. Nothing sneaky.

Personally, I haven't felt it. Claude Code is working fine for me. Nothing that stood out as worse.

But people in the comments told a different story. Context loss, shallow answers, more hallucinations. Real frustration from people who rely on this daily.

So I dug around a bit.

Turns out there's a likely explanation. Anthropic quietly shifted Opus 4.6 to Adaptive Thinking mode with low effort as the default. That means the model decides how much to think per query, and it's been choosing... not much.

The good news is there's a workaround. If you've been feeling a difference, try adding this to your Claude Code settings:

{
  "env": {
    "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1",
    "MAX_THINKING_TOKENS": "128000"
  }
}

Also pair it with "effortLevel": "high" in the same settings.json.

This forces a fixed high-reasoning budget on every turn instead of letting the model cut corners. People are reporting it brings performance back to where it was a few weeks ago.

Why this matters beyond the fix

It's very easy to ship a great product, get people used to a high standard, and then quietly lower the bar. No conspiracy needed. That's just how products work.

Anthropic built a lot of trust on the back of Claude being the thoughtful, reliable one. That reputation is worth protecting. Hope they course correct fast if this is real.

If you try the fix, drop a comment. Curious whether it actually works across different setups.

DEV Community: Zivdev

Is Claude Getting Worse? What I Found After Digging Around