Detailed Analysis
User complaints circulating on Reddit and developer forums in early 2026 reflect a broader pattern of perceived performance degradation in Anthropic's Claude Code — specifically, the observation that earlier versions of the model outperformed its current iteration in complex engineering workflows. The sentiment, captured in the post's blunt framing, aligns with detailed empirical work by Stella Laurenzo, Director of AI at AMD, who analyzed 6,852 Claude Code sessions, 17,871 thinking blocks, and 234,760 tool calls from January through March 2026. Her findings indicated that Claude was reading less code prior to editing, terminating tasks earlier, looping more frequently, and requiring more human corrections on complex tasks — with post-regression prompt language shifting measurably from collaborative to corrective in character.
The evidence, however, points not to alterations of the underlying model weights but to a series of runtime-layer changes implemented by Anthropic in early 2026. On February 9, an "adaptive thinking" feature was introduced, allowing the model to self-determine reasoning depth — a change that caused it to skip deeper reasoning on tasks that appeared simple but were not. Then, on March 3, the default reasoning effort was quietly downgraded from high to medium, producing shallower thinking patterns and more frequent self-corrections visible in outputs (phrases like "oh wait" and "actually" increased in frequency). These runtime modifications, rather than any deliberate degradation of model weights, appear responsible for the functional gap users experience between earlier and current Claude Code behavior. BridgeBench data corroborated some of this concern, showing Claude Opus 4.6 dropping approximately 15 percentage points in accuracy, though that decline may reflect silent compute reallocation toward newer models like 4.7 rather than direct model changes.
Anthropic has publicly denied deliberate downgrades and has recommended that users disable adaptive thinking via the environment variable `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1`, acknowledging the tension without fully resolving user concerns. The company's opacity around telemetry and runtime behavior has become a central criticism — researchers and developers note that Anthropic's own "model diffing" research, which seeks to detect behavioral shifts between model versions, has not been applied publicly to this situation, leaving users without authoritative tools to verify what changed and when. This opacity transforms what might be routine infrastructure optimization into a trust problem, particularly for professional users who have integrated Claude Code deeply into engineering pipelines and who notice even subtle behavioral drift.
The episode reflects a recurring dynamic in the large language model industry, where rapid iteration cycles, opaque deployment practices, and rising user expectations combine to generate "nerf" discourse — complaints that a model has been deliberately weakened, often following a period of exceptional performance. Similar cycles have played out around OpenAI's GPT models. What distinguishes the Claude Code situation is the availability of unusually rigorous session-level data supporting user intuitions, which elevates the conversation above pure vibes-based complaint. The broader implication is that as AI systems become embedded in professional workflows, the distinction between model capability and runtime configuration becomes critically important — and that companies like Anthropic face mounting pressure to provide transparency mechanisms that allow users to understand and audit the actual conditions under which their AI tools are operating, not merely the theoretical capabilities of the underlying model.
Read original article →