Claude Code's Thinking Depth Dropped 67%. Anthropic Stayed Quiet Until Someone Had the Logs.

Claude Code's Thinking Depth Dropped 67%. Anthropic Stayed Quiet Until Someone Had the Logs.

9 min readApril 13, 2026

Something changed with Claude Code around February and a lot of developers felt it. Not a crash. Not a visible error. Just a quiet shift in how the tool approached work. Edits landing without the model reading the file first. More guessing, less research. The kind of regression you blame on yourself before you blame on the tool.

Then on April 2, Stella Laurenzo filed GitHub issue #42796. Laurenzo isn't a random user. She's the Senior Director of AI at AMD, running advanced multi-agent systems across large projects with 50+ concurrent Claude Code sessions doing systems programming in C and MLIR. And she brought receipts.

What the data actually shows

Laurenzo's analysis covered 6,852 session JSONL files from four projects. Not a sample. The whole thing. Inside those sessions: 17,871 thinking blocks, 234,760 tool invocations, and over 18,000 user prompts. The timeframe ran from late January through early April 2026.

Here's what she found:

MetricBefore (Jan-Feb)After (Mar-Apr)Change
Read-to-edit ratio6.6 reads per edit2.0 reads per edit-70%
Stop hook violations~0173 in 17 daysSpiked
User frustration indicators5.8% of prompts9.8% of prompts+69%
Estimated thinking depthBaseline~33% of baseline-67%

That read-to-edit ratio tells the real story. In the good period, Claude's workflow was: read the target file, read related files, grep for usages across the codebase, read headers and tests, then make a precise edit. After the change, it started jumping straight to edits with minimal context gathering.

If you've used Claude Code for anything non-trivial, you know exactly what this feels like. It went from a careful engineer to an intern who skims the spec and starts typing.

What Anthropic actually changed

Boris Cherny from the Claude Code team eventually responded on the Hacker News thread. Two changes happened in close succession:

February 9: Adaptive thinking launched. Opus 4.6 got a new feature where the model decides its own thinking duration instead of receiving a fixed budget. Anthropic described this as working "better overall" but acknowledged it sometimes under-allocates reasoning on specific turns.

March 3: Default effort level dropped to medium. The effort setting went from high to 85/100. Anthropic called this "a sweet spot on the intelligence-latency curve." Faster responses, less deep thinking.

On top of both changes, a thinking redaction header (redact-thinking-2026-02-12) rolled out starting March 5 and reached 100% coverage by March 12. This hid Claude's reasoning from the user interface entirely.

The measurement problem

Part of the 67% figure came from Claude analyzing its own session logs and finding no visible thinking blocks. Boris pointed out that the redaction header obscured them, meaning "the measurement that was broken, not the model itself."

Fair point. But it doesn't explain the behavioral data. The read-to-edit ratio dropping from 6.6 to 2.0 has nothing to do with whether thinking blocks are visible. That's measuring actual tool calls. Claude was reading less before editing. That's not a measurement artifact.

The stop hook violations jumping from zero to 173 in 17 days isn't about thinking visibility either. That's the model violating explicit constraints, something it rarely did before the changes.

So the 67% number might overstate the thinking depth decline specifically. But the quality regression in complex engineering workflows? That part is backed by behavioral data that doesn't depend on thinking block access.

Why this hit power users hardest

The adaptive thinking change works fine for typical single-turn coding tasks. Ask Claude to write a function, it thinks enough, gives you an answer. The problem shows up in long-running, multi-file engineering sessions where the model needs to build up context across many turns.

Laurenzo was running 50+ concurrent sessions doing systems programming. At that scale, a model that occasionally skips reasoning on a turn doesn't just produce a mediocre answer. It produces a confidently wrong one. Her team found instances where Claude fabricated GitHub commits, invented fake package names, and cited incorrect API versions, all on turns where zero reasoning tokens were emitted.

This is the specific bug Boris acknowledged: adaptive thinking sometimes completely skips reasoning, causing hallucinations. Not reduced thinking. Zero thinking. The model just guesses.

Theo breaks down the regression

Theo from t3.gg walked through the full timeline, the data, and what it means for developers relying on Claude Code for production work:

Boris Cherny on what happens after coding is solved

For context on how the Claude Code team thinks about these tradeoffs, Boris Cherny (Head of Claude Code) sat down with Lenny's Podcast. Worth watching to understand the thinking behind adaptive reasoning and effort levels:

The response pattern

What frustrated developers most wasn't the regression itself. Models change. Bugs happen. It was the sequence: users report problems, the default response is "adjust your prompts" or "check your settings," nothing moves until someone produces documentation detailed enough that dismissing it looks bad.

Anthropic has been through this before. In August-September 2025, three infrastructure bugs degraded Claude's response quality for weeks. Their postmortem admitted their evaluations "simply didn't capture the degradation users were reporting." They pledged to build better detection. Eight months later, users are producing the detection that Anthropic's own systems missed.

The story got picked up by The Register, PC Gamer, InfoWorld, and TechRadar. A community post about Claude Code's decline accumulated over 1,060 upvotes. What could have been routine grumbling turned into a genuine trust problem.

The workarounds that exist right now

If you're hitting this, there are concrete fixes. Multiple users confirmed these make "a huge amount of difference":

  • Run /effort high or /effort max in your session to override the medium default.
  • Set CLAUDE_CODE_EFFORT_LEVEL=max as an environment variable so every session starts at full reasoning.
  • Set CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 to force fixed reasoning budgets. This targets the hallucination bug where reasoning gets skipped entirely.
  • Add showThinkingSummaries: true to your settings.json to see what the model is actually reasoning about.

All of these are official configuration options, not hacks. The frustrating part: users had to discover them through community discussion rather than official communication.

What this means for AI coding tools broadly

Claude isn't the only one. IEEE Spectrum reported that AI coding assistants across the board are regressing. A CodeRabbit study found AI-co-authored pull requests have 1.7x more issues than human-authored ones. OpenAI's GPT-4 to GPT-5.x transition broke workflows too.

AI labs optimize for benchmarks, latency, and cost. Developers optimize for "does this tool do what it did last week." Those goals collide more often than anyone admits.

If you depend on AI coding tools for real work, start logging your sessions. Track your own metrics. The model under your tool can change without warning, and the vendor's evals probably won't catch what you notice.

Where Admix fits

This is where AI aggregators earn their keep. When Claude regresses, you want to be able to throw the same prompt at GPT-5.4 or Gemini 3.1 Pro and compare. Admix lets you do that side by side without juggling separate subscriptions.

The race between labs means quality fluctuates constantly. Paying one price for access to all of them turns a regression in one model from a crisis into a five-minute switch.

FAQ

Did Claude Code actually get 67% dumber?

The 67% figure specifically measures thinking depth decline, not overall intelligence. Part of that measurement was affected by thinking redaction hiding the reasoning blocks. The behavioral data (fewer reads before edits, more hook violations) confirms a real quality regression in complex workflows, but the exact percentage is debatable.

Is this fixed now?

Anthropic acknowledged the adaptive thinking bug and the effort level default. Setting /effort max or CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 addresses the worst symptoms. Whether the underlying defaults will change is unclear.

Should I stop using Claude Code?

Not necessarily. With the workarounds applied, many users report quality returning to previous levels. The issue is specifically with the default settings for complex, multi-file engineering work. For simpler tasks, the default effort level works fine.

Who filed the original issue?

Stella Laurenzo, Senior Director of AI at AMD. She was running advanced multi-agent systems with 50+ concurrent Claude Code sessions for systems programming in C and MLIR.

Sources

Further Reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.

Start free on Admix

Related articles