Smoke evaluation data from 2026-11-09 to 2026-11-15 shows GPT-5.5 rising from 78.74 on the first day to 88.75 on the last, with a trend value of +10, making it the model with the largest increase this week.
Score Trajectories of Rising Models
GPT-5.5 averaged 75.3 points with volatility of 37.7, and its last-day score of 88.75 marked a clear improvement over the first day. Claude Sonnet 4.6 rose from 91 on the first day to 95.19 on the last, with a trend of +4.2 and an average of 78.9. GPT-6 Luna rose from 90.24 on the first day to 95.19 on the last, with a trend of +5 and an average of 81.9. Claude Opus 4.7 rose from 78.74 on the first day to 81.44 on the last, with a trend of +2.7 and an average of 81.2. GPT-o3 rose from 92.49 on the first day to 94.31 on the last, with a trend of +1.8 and an average of 75.5.
Score Trajectories of Declining Models
Gemini 2.5 Pro fell from 82.52 on the first day to 51.63 on the last, with a trend of -30.9 and an average of 57.4. Gemini 3.1 Pro fell from 78.74 on the first day to 53.94 on the last, with a trend of -24.8 and an average of 65.4. GLM-4.6 fell from 36.25 on the first day to 0 on the last, with a trend of -36.3 and an average of 25.5. Doubao Pro fell from 83.49 on the first day to 65.38 on the last, with a trend of -18.1 and an average of 71.5. DeepSeek V4 Pro fell from 68 on the first day to 49.44 on the last, with a trend of -18.6 and an average of 53.6. GPT-6 Sol fell from 92.49 on the first day to 90.69 on the last, with a trend of -1.8 and an average of 78.7. Grok 4 fell from 91.33 on the first day to 82.7 on the last, with a trend of -8.6 and an average of 71.2. GPT-6 Astra fell from 87.99 on the first day to 84.7 on the last, with a trend of -3.3 and an average of 80. GPT-6.1 Sol fell from 87.99 on the first day to 84.7 on the last, with a trend of -3.3 and an average of 82.1.
Stability Analysis of High-Volatility Models
GLM-4.6 showed volatility of 53.7 points, far higher than other models, indicating a large standard deviation in daily scores and the lowest consistency. Grok 4 showed volatility of 42.7 points, GPT-5.5 37.7 points, Gemini 2.5 Pro 37.5 points, GPT-o3 36.7 points, Claude Sonnet 4.6 35.3 points, GPT-6 Luna 34.8 points, and GPT-6 Astra 36.2 points. GPT-6.1 Sol showed volatility of 20.4 points, Doubao Pro 25.9 points, DeepSeek V4 Pro 25.2 points, Claude Opus 4.7 28.5 points, Gemini 3.1 Pro 28.5 points, and GPT-6 Sol 26.6 points. Qwen3 Max showed volatility of 0 points, remaining at 0 on the last day.
Integrity Rating Change Signals
Grok 4's integrity rating shifted from pass to warn and then recovered to pass, with one warn record during the period. GLM-4.6's integrity rating shifted from pass to fail, then to warn, followed by pass and then warn again, changing multiple times. Other models maintained a pass integrity rating with no recorded changes.
Causal Mechanism Analysis
GLM-4.6's last-day score of 0 with volatility of 53.7 points most likely stems from losing points on multiple of the 10 daily questions, widening the standard deviation. Gemini 2.5 Pro's trend of -30.9 points, with the gap between its first-day 82.52 and last-day 51.63, directly reflects a decline in material-constraint dimension scores. GPT-5.5's trend of +10 points and average of 75.3 indicate a gradual improvement in code-execution dimension scores over seven consecutive days. When Grok 4's integrity rating warn appeared, its volatility of 42.7 points coexisted, indicating instability in the engineering-judgment dimension.
Practical Implications for Users
Teams focused on code execution choosing GPT-5.5 can refer to its last-day score of 88.75 and trend of +10, suitable for scenarios requiring stable output over multiple consecutive days. For scenarios sensitive to material faithfulness, Gemini 2.5 Pro's last-day score of 51.63 and trend of -30.9 suggest additional verification of output accuracy is needed. Developers relying on consistency should avoid GLM-4.6, whose volatility of 53.7 points and multiple integrity rating changes directly increase integration risk. For enterprise model selection, GPT-6 Luna's average of 81.9 and trend of +5 can serve as a reference benchmark.
Strategic Judgment
Based on its trend of +10 points and last-day score of 88.75, GPT-5.5 appears undervalued in this week's Smoke evaluation. Gemini 2.5 Pro and Gemini 3.1 Pro both showed declines exceeding 24 points, with averages of 57.4 and 65.4 respectively, a signal worth verifying next period. GLM-4.6's path from pass to fail to warn, together with its volatility of 53.7 points, forms a high-risk combination. Claude Sonnet 4.6 and GPT-6 Luna both reached 95.19 on the last day, with trends of +4.2 and +5 respectively, and can serve as parallel options for similar scenarios.
This week's data chain is clear: GPT-5.5, Claude Sonnet 4.6, GPT-6 Luna, Claude Opus 4.7, and GPT-o3 show positive trends, while the Gemini series, GLM-4.6, Doubao Pro, and DeepSeek V4 Pro show significant negative trends. High-volatility models are concentrated in GLM-4.6 and Grok 4, and integrity rating changes appeared only in Grok 4 and GLM-4.6. All judgments derive directly from the first-day, last-day, average, trend, and volatility values from 2026-11-09 to 2026-11-15.
Data source: YZ Index | Run #427 | View raw data
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接