Kaggle FACTS Grounding: leaderboard
Each item supplies a source document of up to about 32,000 tokens, an instruction to rely only on that document, and a request such as a summary, a rewrite or a question. Responses that fail to satisfy the request are disqualified, then several frontier model judges decide whether the rest are fully supported by the document and their verdicts are averaged. This board is the version 2 leaderboard, higher is better.
Metric: Score (%). Source: www.kaggle.com. Saturation forecast: Around January 2027. 52 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Sol | 85.22 | #14 |
| 2 | GPT-6.1 Sol | 82.42 | #7 |
| 3 | GPT-6 Astra | 81.98 | #6 |
| 4 | Gemma 4 26B A4B | 80.88 | #358 |
| 5 | Gemma 4 31B | 80.67 | #200 |
| 6 | Gemini 4 Argon | 78.63 | #8 |
| 7 | GPT-5.2 | 76.2 | #110 |
| 8 | GPT-5.5 | 76.19 | #31 |
| 9 | Gemini 3.7 Flash | 75.54 | #23 |
| 10 | Grok 4.5 | 74.52 | #48 |
| 11 | Gemini 2.5 Pro | 74.29 | #152 |
| 12 | Gemini 3.5 Flash | 73.24 | #60 |
| 13 | Gemini 3.8 Flash | 73.09 | #21 |
| 14 | Gemini 3.6 Flash | 72.51 | #51 |
| 15 | GPT-5.6 Terra | 71.21 | #41 |
No result here: #9 Claude Fable 5.1, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=kaggle-facts-grounding · How It Works · Data refreshed daily, snapshot 2026-10-20.