Question 1

Did Claude actually get dumber in March 2026?

Accepted Answer

Yes, but the underlying model didn't change. Anthropic shipped three silent default changes to the scaffolding around Opus 4.6 between February 9 and March 3, 2026: adaptive thinking became the default, a redact-thinking header started hiding internal reasoning from session logs, and the default effort level dropped from high to medium. The cumulative effect for power users was 43 mid-task bail-outs per day, visible reasoning length down 67%, and self-contradiction roughly tripled.

Question 2

What changed between February and March 2026 that broke Claude?

Accepted Answer

Three product changes confirmed by Anthropic engineer Boris Cherny on Hacker News. February 9: adaptive thinking became the default, meaning the model now decides per-turn how much reasoning to spend. February 12: a redact-thinking header started hiding internal reasoning from session logs. March 3: the default effort level dropped from high to medium. None were weight changes. The underlying model didn't get dumber. The harness around it got retuned, and prompts dialed in for January-era defaults were suddenly talking to a different operating environment.

Question 3

How do I fix Claude when it stops mid-task or keeps asking for permission?

Accepted Answer

Set CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 in your environment. This forces a fixed reasoning budget per turn and is the highest-leverage single fix. It addresses the asking-for-permission pattern, the skipped file reads, and the confident fabrication on complex tasks all at once. Combine with /effort high at the start of complex sessions, or /effort max for architecture decisions, hard debugging, or multi-file refactors.

Question 4

How do I disable Claude's adaptive thinking?

Accepted Answer

Set the environment variable CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1. Adaptive thinking became the default on February 9, 2026, and routes every turn through a classifier that asks simple-or-complex. The classifier was trained on general usage and sometimes routes a complex multi-file refactor as text replacement, dropping reasoning to near zero. Zero-reasoning turns are when the model fabricates with confidence. Disabling adaptive thinking forces a fixed reasoning budget and stops the zero-reasoning failure mode.

Question 5

What is the actual usable context window for Claude Code?

Accepted Answer

Treat your usable context as roughly 400K tokens, not the advertised 1M. Documented bug reports show quality dropping at 20% utilization, the model compressing earlier context at 40%, and Claude itself recommending a fresh session at 48%. A separate report documented Claude Code firing its compaction routine at ~76K tokens on a 1M session, discarding history with 924K tokens of headroom still free. Use the Document and Clear pattern: have the model summarize progress to a markdown file, run /clear, and restart with the markdown in context.

Question 6

Should I use Opus or Sonnet for coding now?

Accepted Answer

Sonnet for most production tasks. The gap between Opus and Sonnet is smaller in 4.6 than in any prior generation: SWE-bench Verified shows Sonnet 4.6 at 79.6% versus Opus 4.6 at 80.8%, a rounding error on most coding work. Reserve Opus with /effort max and adaptive thinking disabled for genuinely complex multi-file engineering, long-horizon planning, or architecture decisions. Use Haiku 4.5 for documentation, formatting, and simple queries.

Claude actually did get dumber, but you can…