Claude telemetry shows nerf
An AMD senior AI director posted telemetry claiming Claude has been 'nerfed'—showing fewer 'thinking' tokens, more retries and performance drops when load increases across about 7,000 sessions. The analysis and calls for '/effort max' workarounds were shared on social telemetry posts that include session data observations (x.com).
An Advanced Micro Devices AI director said Anthropic’s Claude Code has become less reliable on complex engineering work, citing telemetry from 6,852 sessions logged between January 30 and April 1. (github.com) Stella Laurenzo posted the analysis in Anthropic’s public GitHub issue tracker on April 2, saying her team reviewed 17,871 thinking blocks and 234,760 tool calls from a “very consistent, high complexity work environment.” The issue says the quality drop began after February updates and that Claude “cannot be trusted to perform complex engineering.” (github.com) Her logs tie the shift to a rollout labeled “redact-thinking-2026-02-12.” In the issue, visible thinking fell from 100 percent before March 5 to 41.6 percent by March 8, while redacted thinking rose to 58.4 percent. (github.com) Claude’s “thinking” feature is the extra scratch-pad time the model uses before it answers, like letting a developer sketch a plan before editing code. Anthropic introduced that mode with Claude 3.7 Sonnet on February 24, 2025, and said developers could trade speed and cost for better answers by giving the model more room to think. (anthropic.com) Anthropic’s current developer docs make the same tradeoff explicit. The company says the effort setting controls how many tokens Claude spends, and that lower effort can mean fewer tool calls and less thorough work, while “high” is the default and “max” is reserved for the highest capability. (platform.claude.com) That is why Laurenzo’s workaround advice centered on forcing higher effort. Posts amplifying her analysis on X urged users to try “/effort max,” a command that maps to Anthropic’s documented setting for deeper reasoning and higher token use. (platform.claude.com) The dispute lands after Anthropic publicly acknowledged a different quality problem in 2025. In a September 17, 2025 postmortem, the company said three infrastructure bugs had intermittently degraded responses and added, “We never reduce model quality due to demand, time of day, or server load.” (anthropic.com) Anthropic has not published a matching postmortem for Laurenzo’s April 2026 complaint on the sources reviewed here. The public GitHub issue was later marked closed, but the issue text remains visible and continues to circulate in developer discussions and media coverage. (github.com) The thread running through both Anthropic’s docs and Laurenzo’s logs is simple: when Claude spends fewer tokens thinking, it may move faster, but power users say it also makes more shallow edits. That is the claim now under scrutiny as developers test whether “max” effort restores the behavior they were getting in January. (platform.claude.com)
