Why Claude Code Is Getting Dumber: Data‑Driven Dive into AI Programming Decline

1. Measured Drop in Thought Depth

Stella Laurenzo collected 6,852 Claude Code sessions covering 17,871 reasoning blocks and 234,760 tool calls from late January to early April 2026 across four real projects. The average reasoning length fell from 2,200 characters to 720 characters (‑67%) and later to 560 characters (‑75%). This shift equates to moving from a “thoughtful expert” to an “intern doing chores”.

2. Manifestations of the Decline

Simplest Fix Syndrome

When the model outputs “Simplest Fix”, it chooses the lowest‑cost reasoning path instead of evaluating multiple solutions.

Skipping Reading, Rewriting Whole Files

Previously the model read code on average 6.6 times per edit, making precise adjustments.

Now it reads only twice and often rewrites entire files.

Full‑file recreation rose from 4.9 % to 11.1 % of edits.

Premature reasoning interruptions and “permission‑seeking” shortcuts jumped from zero before March 8 to an average of ten per day.

Admitting Laziness

After users correct it, Claude explicitly says “you’re right, this is too lazy” or “I was too hasty”, indicating the model only detects the error after output.

3. Is This Unique to Claude?

Historical data shows Claude Opus 4.1 suffered a similar drop in August‑September 2025, which Anthropic publicly rolled back. OpenAI’s GPT‑4o has faced user accusations of “getting lazy” but the company denies systematic degradation. Some domestic models show occasional quality spikes during peak load.

4. Root Causes

The primary driver is the high inference cost of large models. To preserve margins during peak demand, providers apply:

Dynamic quantization (e.g., 1.58‑bit ternary quantization replaces FP16, causing sharp accuracy loss).

Inference truncation (shortening reasoning chains to cut token usage).

Routing downgrade (steering requests to smaller model versions).

These optimizations improve gross margin on paper but manifest to users as “downgraded intelligence”.

5. Recommendations for Developers

Adopt a multi‑model backup strategy: keep Claude, GPT‑4o, DeepSeek, etc., ready; Codex usage spiked ten‑fold during Claude’s outage.

Monitor the “thought depth” metric; signs such as “Simplest Fix”, low read counts, or full‑file rewrites suggest lazy behavior. Enabling an “Extended Thinking” mode can mitigate.

Schedule heavy engineering tasks outside peak windows (e.g., Pacific night or before noon Beijing time) to avoid degraded service.

Consider local fallback models (DeepSeek, Llama) for critical workloads; they may be less capable but offer stable performance.

6. Closing Thoughts

Anthropic’s public statement that it will not intentionally lower model quality contrasts with the reality that commercial pressure can force trade‑offs, turning a leading programming AI into a costly “artificial idiot”. Developers should remember that tools augment, not replace, human coding skill.