Stella Laurenzo's AMD Telemetry Analysis of Claude Opus 4.6 Regression
In April 2026, AMD's Senior Director of AI filed a GitHub issue analyzing 6,852 Claude Code sessions and 17,871 thinking blocks, quantifying a 73% collapse in median thinking output and other regressions that pushed AMD's engineering team to a competitor.
On April 2, 2026, Stella Laurenzo, Senior Director of AMD's AI group, filed GitHub issue #42796 documenting a measurable regression in Claude Opus 4.6. Her dataset covered 6,852 Claude Code sessions, 17,871 thinking blocks, and 234,760 tool calls. Key findings: - Median thinking block length collapsed about 73%, from roughly 2,200 characters to 600. - File-reading-before-editing was skipped 33.7% of the time, up from a 6% baseline. - Full-file rewrites occurred roughly twice as often (a wasteful pattern when a targeted edit would do). - Mid-task abandonment rose from previously-zero rates. - Users needed about 12 times more interruptions to keep the model on track. - Hallucination of commit SHAs, package names, and API versions increased significantly. - The word 'simplest' appeared in output about three times more often, suggesting the model was defaulting to surface-level solutions. AMD's engineering team switched to a competitor model after the analysis. The viral BridgeBench claim of 83.6% to 68.3% accuracy drop was methodologically weak but directionally consistent with Laurenzo's telemetry. The episode is notable as one of the few cases where a major customer published quantitative evidence of silent model degradation rather than relying on subjective complaints. See Claude Opus 4.6 Silent Degradation Timeline (February-March 2026) for the underlying changes that produced these effects.