GPT-6 Luna Tops the List with 82.77 Points: 2026-11-17 Smoke Quick-Test Data Brief

The 2026-11-17 YZ Index Smoke quick test covered 14 models, with GPT-6 Luna ranking first for the day at 82.77 points. The Smoke test is a daily 10-question quick assessment suited to spotting short-term signals; it is not equivalent to the conclusions of the weekly Full leaderboard.

This Smoke evaluation covered only two main-leaderboard dimensions: code execution and material constraints. The main leaderboard formula is 0.55 × code execution + 0.45 × material constraints. Because the daily sample size is small, single-day scores are better treated as monitoring signals rather than as long-term conclusions about model capability.

Rankings for the Day

RankModelMain LeaderboardCode ExecutionMaterial ConstraintsIntegrity
#1GPT-6 Luna82.7799.162.8pass
#2GPT-5.581.5899.859.3pass
#3Claude Opus 4.778.0392.760.1pass
#4GPT-6.1 Sol76.499.847.8pass
#5GPT-6 Sol75.8499.147.4pass
#6Claude Sonnet 4.670.2683.853.7pass
#7GPT-6 Astra65.6274.854.4pass
#8GPT-o365.2766.364pass
#9Doubao Pro59.8359.660.1pass
#10Gemini 3.1 Pro56.2948.565.8pass
#11Grok 449.0950.847pass
#12GLM-4.643.362565.8pass
#13Gemini 2.5 Pro42.3328.659.1pass
#14DeepSeek V4 Pro39.272556.7pass

Interpreting the Data

Looking at the score structure, the leading models mostly show complementary pairings of code execution and material constraints. GPT-6 Luna posted 99.1 in code execution and 62.8 in material constraints for a main-leaderboard score of 82.77; GPT-5.5 took second with 99.8 in code execution and 59.3 in material constraints; and Claude Opus 4.7 placed third with 92.7 in code execution and 60.1 in material constraints. Such pairings keep overall scores in the lead. GPT-6.1 Sol scored 99.8 in code execution but 47.8 in material constraints, while GPT-6 Sol scored 99.1 in code execution and 47.4 in material constraints, showing a pattern of strong code execution paired with relatively weak material-constraint performance.

Among models with notable changes, Grok 4 fell 32.9 points on the main leaderboard, with code execution down 25.2 and material constraints down 42.3; GPT-6 Sol fell 16.6 points on the main leaderboard, with material constraints down 41.9; GPT-6 Astra fell 13.1 points on the main leaderboard, with material constraints down 34.9; Doubao Pro rose 19 points on the main leaderboard, with code execution up 38; and DeepSeek V4 Pro fell 11.9 points on the main leaderboard, with material constraints down 32.6. These single-day fluctuations may stem from question sampling variance, or they may reflect genuine differences in how models perform under specific constraints; follow-up runs using the same methodology are needed to confirm.

As a small-sample, single-day signal, Smoke currently provides only same-day observations, and no anomalous signals were observed; the interpretation remains restrained to reflect the actual characteristics of the score structure.

Key Changes

  • Grok 4: main leaderboard down 32.9 points, code execution down 25.2, material constraints down 42.3
  • Doubao Pro: main leaderboard up 19 points, code execution up 38
  • GPT-6 Sol: main leaderboard down 16.6 points, material constraints down 41.9
  • GPT-6 Astra: main leaderboard down 13.1 points, material constraints down 34.9
  • DeepSeek V4 Pro: main leaderboard down 11.9 points, code execution up 5, material constraints down 32.6

Signals to Watch

  • No publishable anomalous signals were retained for this run.

When reading this kind of Smoke brief, the focus should be on two questions: first, whether a model exposes the same type of weakness on consecutive days; and second, whether its integrity rating moves from pass to warn or fail. Large single-day changes in execution or constraint scores may come from question sampling, or they may be early signals of genuine degradation, and follow-up runs are needed to confirm.


Data source: YZ Index | Run #431 | View raw data

This article is from Winzheng Index blog, translated in full by Winzheng (winzheng.com). Click here to view the original When republishing the translation, please credit the source. Thank you!