GPT-6 Luna’s main leaderboard score fell from 86.95 to 67.17 in today’s Smoke evaluation, a single-day drop of 19.8 points.
Changes in Core Dimension Scores
The code execution dimension fell from 100.00 yesterday to 75.00, down 25 points; the material constraints dimension fell from 71.00 to 57.60, down 13.4 points; the engineering judgment dimension remained at 100.00; and task expression fell from 87.50 to 75.00, down 12.5 points. The integrity rating remained pass.
Possible Mechanisms Behind the Score Decline
The Smoke evaluation has only 10 questions per day, with two questions corresponding to one dimension, so randomness in question sampling may cause single-day score fluctuations. The code execution dimension fell by 25 points this time, and material constraints fell by 13.4 points; together they contributed the 19.8-point drop on the main leaderboard. The engineering judgment dimension showed no change, indicating that the model remains stable in that side leaderboard capability.
In terms of dimension composition, code execution and material constraints are the only two auditable dimensions on the main leaderboard. The simultaneous decline of both dimensions this time indicates that the model incurred more deductions when handling specific code tasks and material fidelity constraints. The task expression side leaderboard fell by 12.5 points, further showing that the model’s output fluctuated in following task instructions.
Specific Implications for Users
Teams that rely heavily on code execution face higher risk in today’s evaluation. The code execution dimension fell from 100.00 to 75.00, meaning the model made more errors when generating executable code or completing programming tasks. Scenarios that depend on material constraints are also affected: material constraints fell from 71.00 to 57.60, indicating that the model deviated more when strictly following input materials.
The engineering judgment dimension remained at 100.00, with relatively little impact on developers who need engineering judgment assistance. The integrity rating remained pass, and the model showed no problems in basic compliance.
Strategic Judgment
The 19.8-point drop on the main leaderboard this time was mainly driven by the 25-point decline in the code execution dimension. The number of questions in a single-day Smoke evaluation is limited, and the current data are insufficient to distinguish between fluctuations from question sampling and genuine model degradation. The zero change in the engineering judgment dimension shows that some model capabilities remain stable.
For enterprise model selection, it is necessary to observe in subsequent multi-day evaluations whether the two main leaderboard dimensions, code execution and material constraints, remain below yesterday’s levels. If the scores in both dimensions recover in the next evaluation period, the decline is more likely due to question randomness; if they continue to decline, the model’s suitability for code and material-constraint scenarios needs to be reassessed.
Based on the existing score comparison, the stability of GPT-6 Luna in code execution scenarios warrants verification in the next period. The 13.4-point decline in the material constraints dimension also needs continued tracking to confirm its impact on applications that depend on material fidelity.
Data source: YZ Index | Run #401 | View raw data
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接