Significant quality regression — feels like a lower-tier model is being served

Since the recent usage/billing bug and subsequent changes, Claude Code (Opus, v2.1.92, Max subscription) feels substantially nerfed. I'm now having to use OpenAI Codex as a second opinion for plan reviews, code reviews, and test design — work that Claude Code used to handle confidently on its own.

What's changed

• No deep thinking — jumps to surface-level fixes instead of reasoning through the problem. Previously considered root causes and edge cases before acting; now tries the first thing that comes to mind.
• Lost logical coherence — contradicts its own findings. e.g., confirms data values are valid integers, then "fixes" the issue by adding null checks. Doesn't connect the dots.
• 90% of deliveries are unfinished — claims a task is "done" without actually verifying. I have to ask it to double-check almost every time. When I do, it finds the issue was never actually fixed.
• Delegates verification back to me — ~70% of the time it says "try it and let me know if it works" instead of testing it itself. I'm paying for an AI engineer, not becoming QA for one.
• Blame-shifts when caught — when I point out something isn't done, the response is "sorry" followed by blaming an unrelated issue, rather than investigating what actually went wrong.
• Shallow context — forgets constraints from 2 messages ago. Doesn't hold the full picture of what it's working on.
• Needs constant babysitting — I now delegate plan review and code review to Codex CLI to catch what Claude misses. This defeats the purpose of paying for a Max subscription.

The role reversal problem

The user should be directing work and reviewing results. Instead, Claude delivers half-finished work, asks the user to verify, and the user ends up debugging Claude's output. The delegation has flipped — I feel like I've been assigned to check whether Claude did its job, rather than Claude checking its own work before presenting it.

Before: Felt like working with a senior engineer — would think through the problem, verify the fix, catch edge cases, and push back when something didn't make sense.

Now: Feels like managing an intern who skims the ticket, does the minimum, says "done", and waits for you to find what's broken.

Question

Is there silent model routing or throttling happening on Max subscriptions after the usage bug? If Opus is being substituted with a lighter model under load, that would explain the experience. Transparency here would be appreciated — I'd rather know I'm getting Sonnet than be told it's Opus when it clearly isn't performing like it.

Environment

• Claude Code v2.1.92
• macOS
• Max subscription
• Opus selected