GPT-5.5 Material Constraint Score Drops 15.2 Points, Code Execution Rises 30 Points, Main Leaderboard Up 9.7

In today's Smoke evaluation, GPT-5.5's material constraint score dropped from 65.80 to 50.60, a decline of 15.2 points; its code execution score rose from 45.00 to 75.00, an increase of 30 points; and its main leaderboard score rose from 54.36 to 64.02, an increase of 9.7 points.

Score Change Facts

The engineering judgment score rose from 75.00 to 100.00, an increase of 25 points; the task expression score dropped from 90.00 to 81.70, a decline of 8.3 points. The integrity rating remains "pass."

Possible Cause Analysis

The Smoke evaluation includes only 10 questions per day, with 2 questions per dimension. Given the small sample size, single-day fluctuations are normal. The opposing movements in material constraint and code execution scores most likely stem from differences in the day's question distribution rather than structural degradation of model capabilities. The perfect engineering judgment score and the decline in task expression similarly point to shifts in question focus.

Implications for Users

Enterprises with material-constraint-heavy scenarios should note the 50.60 score for the day. Developers relying on GPT-5.5 for long-document or citation tasks should add manual verification steps. Teams focused on code execution can take short-term advantage of the 75.00 score, but should monitor multi-day data to confirm stability.

Strategic Assessment

The rise in the main leaderboard score was primarily driven by code execution. The 15.2-point drop in material constraint does not constitute a clear degradation signal given the small sample size. It is recommended to continue tracking the parallel trends of material constraint and code execution in the next round; if the decline persists, further validation will be needed.

The opposing movements of 30 points and 15.2 points between material constraint and code execution in this evaluation confirm that the daily 2-question draw has a greater impact on scores than short-term model changes. The engineering judgment score rising to 100.00 further supports the question-difficulty-distribution explanation.

For material-constraint-dependent scenarios, GPT-5.5's 50.60 score today suggests citation accuracy may have declined, and users should add a second check at critical output stages. The 75.00 code execution score offers temporary convenience for developers needing rapid prototype validation.

Taking the main leaderboard score of 64.02 and all dimension data into account, GPT-5.5's current performance shows no systematic degradation; the single-day drop in material constraint is more likely a result of draw fluctuations. Continuing to track the standard deviation of material constraint scores in future Smoke evaluations will provide a more reliable basis for assessment.

Data source: YZ Index | Run #279 | View raw data