Claude Opus 4.7's main leaderboard score in the Smoke evaluation dropped from 96.99 to 82.95, its Code Execution dimension fell from 100.00 to 69.00, a drop of 31 points.
Score change breakdown
The Material Constraint dimension rose from 93.30 to 100.00, Engineering Judgment rose from 55.60 to 100.00, and Task Expression fell from 95.80 to 78.90. Since the main leaderboard is weighted solely by Code Execution and Material Constraint, the 31-point decline in Code Execution directly pulled down the overall score by 14.04 points.
Possible cause analysis
The Smoke evaluation has only 2 questions per dimension per day. Today's Code Execution dimension had a total score of only 69 across 2 questions, far below yesterday's perfect score. Meanwhile, Material Constraint reached a perfect score in the same period, indicating no systemic degradation in the model's fundamental capabilities. The most likely cause is that today's Code Execution questions differed from yesterday's draw in terms of boundary condition handling or multi-step debugging.
Engineering Judgment (side leaderboard, AI-assisted evaluation) rose to 100 points while Task Expression dropped by 16.9 points, further confirming that fluctuations are concentrated in specific dimensions rather than a global capability decline. The Integrity Rating remains "pass," with no consistency anomalies.
Implications for users
Teams that rely heavily on code execution should allow headroom in daily Smoke scenarios; a single-day score of 69 may increase the failure rate of automated script generation. The perfect score on Material Constraint indicates that scenarios sensitive to document fidelity can still confidently use Claude Opus 4.7.
Developers who depend on continuous multi-turn code debugging should add local validation steps before critical tasks to avoid direct impact from single-day draw fluctuations on delivery.
Strategic judgment
The 14-point drop on the main leaderboard is primarily caused by draw fluctuations in 2 Code Execution questions, not a real model degradation signal. If Code Execution rebounds above 95 in the next Smoke evaluation, today's event can be confirmed as an isolated incident; if it remains below 80, it is necessary to track changes in the composition of the Code Execution dimension.
Current data does not support a downward revision of Claude Opus 4.7's overall capability. It is recommended that enterprises evaluating the model continue to use yesterday's 96.99 baseline, and only add redundancy checks in high-frequency code generation scenarios.
Data source: YZ Index | Run #229 | View raw data
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接