Telemetry claim: Claude 'nerfed' internally
An AMD senior AI director posted telemetry alleging Anthropic reduced Claude’s median ‘thinking’ length from roughly 2,200 to about 600 characters, which the post says increased API requests ~80× and produced more self‑contradictions across ~7,000 sessions. The same post noted a workaround using the '/effort max' prompt. (x.com) (x.com)
A large language model’s “thinking” is the extra internal work it does before answering, and Anthropic’s own docs say more effort usually means more capability. An AMD artificial intelligence director now says Claude Code has been doing far less of that work since late February. (platform.claude.com) (gist.github.com) The claim comes from Stella Laurenzo, who is identified in outside coverage as director of the artificial intelligence group at Advanced Micro Devices, and from a telemetry analysis of 6,852 Claude Code session files, 234,760 tool calls, and 17,871 thinking blocks. A linked write-up says median thinking length fell from about 2,200 characters in the January 30 to February 8 baseline to about 720 in late February and about 600 after March 12. (winbuzzer.com) (gist.github.com) Anthropic’s documentation says the effort setting controls how many tokens Claude spends, and says “high” is the default while “max” is the highest-capability setting. The same docs say lower effort can mean less thinking and fewer tool calls, while higher effort spends more tokens for tougher coding and agent tasks. (platform.claude.com 1) (platform.claude.com 2) That matters in Claude Code because the product is designed to read files, run commands, edit code, and keep working through long tasks on its own. Anthropic’s best-practices guide says these sessions can consume tens of thousands of tokens and that performance degrades as context fills up. (anthropic.com) (code.claude.com) Anthropic has also been pushing newer controls around this behavior. Its February 5 launch post for Claude Opus 4.6 introduced adaptive thinking and new effort controls, and current API docs recommend adaptive thinking plus the effort parameter for Claude Opus 4.6 and Claude Sonnet 4.6. (anthropic.com) (platform.claude.com) Laurenzo’s analysis says the visible thinking redaction rollout ran from March 5 to March 12, moving from 1.5 percent redacted to fully redacted. The same report says the drop in estimated thinking depth started before that rollout, with a 0.971 Pearson correlation between a signature field and visible thinking length used to estimate post-redaction depth. (gist.github.com) The report also says behavior changed after March 8: stop-hook violations rose from zero to 173, user-prompt frustration markers rose from 5.8 percent to 9.8 percent, and sessions with reasoning loops rose from zero to seven. A separate report from WinBuzzer said Laurenzo described Claude Code as reading code three times less before editing and rewriting entire files twice as often. (gist.github.com) (winbuzzer.com) Anthropic’s public docs do not say it lowered the default effort in Claude Code, and the API docs currently say omitting the effort parameter is equivalent to “high.” The same docs also say “max” is available on Claude Mythos Preview, Claude Opus 4.6, and Claude Sonnet 4.6 for the deepest reasoning. (platform.claude.com 1) (platform.claude.com 2) That leaves two things true at once on April 12, 2026: Anthropic officially documents effort as a direct tradeoff between capability and token use, and Laurenzo’s telemetry claims Claude Code has recently been spending much less of that budget unless users push it harder themselves. (platform.claude.com) (gist.github.com)
