Anthropic’s Claude reportedly nerfed
An AMD senior AI director posted an analysis claiming Anthropic’s Claude model has been 'nerfed'—showing fewer 'thinking' tokens, more retries and performance that varies with load, with a suggested workaround of using '/effort max'. The thread circulated on social media and sparked debate about model behaviour and throttling. (x.com 1) (x.com 2)
Claude Code, Anthropic’s tool for editing code and running commands, is under scrutiny after an Advanced Micro Devices director published logs from 6,852 sessions arguing the model got worse after February updates. (github.com) (code.claude.com) The post came from Stella Laurenzo, identified by The Register and WinBuzzer as an Advanced Micro Devices artificial intelligence director, and it said her team analyzed 17,871 thinking blocks and 234,760 tool calls from January through March 2026. (theregister.com) (winbuzzer.com) (github.com) Her GitHub issue said the timeline of “thinking redaction” matched the regression: visible reasoning was 100 percent from January 30 to March 4, then redacted output rose on March 5, March 7, and March 8 as quality metrics fell. (github.com) (theregister.com) To understand the complaint, start with how Claude Code works. Anthropic says the product can read a codebase, edit files, run shell commands, and use tools across a project, so users judge it not just by answers but by whether it researches before it changes code. (code.claude.com) Anthropic’s newer models also use “adaptive thinking,” which means the model decides how much internal reasoning to spend on a request, and an “effort” setting that trades speed and cost against depth. Anthropic’s documentation says the default effort is “high” in the API, while “max” spends more tokens for the strongest capability. (anthropic.com) (platform.claude.com 1) (platform.claude.com 2) Claude Code has its own controls on top of that. Anthropic’s docs say users can switch models in-session, disable adaptive reasoning with `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1`, and use a “max” effort level on supported models, though that setting is slower and costs more. (code.claude.com 1) (code.claude.com 2) (platform.claude.com) Laurenzo’s report argued the tool’s behavior shifted from “research-first” to “edit-first.” The Register said her team saw average code reads fall from 6.6 to 2 by late March, while whole-file rewrites and stop-hook violations increased after March 8. (github.com) (theregister.com) That claim spread because Anthropic had introduced the relevant knobs only weeks earlier. On February 5, 2026, Anthropic announced Claude Opus 4.6 with adaptive thinking and new effort controls, describing them as ways to balance intelligence, speed, and cost. (anthropic.com) Anthropic’s own documentation cuts against the idea that every shorter answer is a secret downgrade. The company says lower effort can mean fewer tool calls and less thinking on simpler problems, and that adaptive mode lets the model decide when deep reasoning is necessary. (platform.claude.com 1) (platform.claude.com 2) The practical takeaway from the debate is narrower than the social-media slogan. If a Claude Code session starts acting fast but shallow on a hard engineering task, Anthropic’s docs point users toward explicit model selection, higher effort, and disabling adaptive thinking before concluding the base model itself has been permanently “nerfed.” (code.claude.com) (platform.claude.com)
