FACTS Grounding v2
We show this table for reference; we do not rank on it.
FACTS Grounding v2 checks whether long-form answers remain supported by supplied documents. Its revised judge ensemble and public/private results remain separate from version 1.
Evaluation results published by Google DeepMind, Google Research, and Kaggle. How Kaggle Benchmarks reruns tests.
Kaggle labels this benchmark Apache 2.0. We retain aggregate measurements with attribution and source links. That declaration does not establish rights to third-party documents, videos, or other underlying task material; no task contents are included. Original article terms.
Grounding score (%) on FACTS Grounding v2 — October 13, 2026
We mirror the published grounding score (%) view for FACTS Grounding v2. GPT-6 Sol has the highest reported value at 85.2%. These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. We do not use these results to rank models overall.
GPT-6 Sol
Kaggle benchmark v2; task revision 3; Evaluated 2026-09-25
GPT-6.1 Sol
Kaggle benchmark v2; task revision 3; Evaluated 2026-10-13
GPT-6 Astra
Kaggle benchmark v2; task revision 3; Evaluated 2026-09-11
55 source rowsReasoningCurrentDisplay onlyUpdated October 13, 2026
Grounding score (%) table (55 source rows)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 SolCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-09-25 | 85.2% | Not reported | Unknown |
| 2 | GPT-6.1 SolCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-10-13 | 82.4% | Not reported | Unknown |
| 3 | GPT-6 AstraCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-09-11 | 82.0% | Not reported | Unknown |
| 4 | Gemma 4 26B A4BCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 80.9% | Not reported | Unknown |
| 5 | Gemma 4 31BCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 80.7% | Not reported | Unknown |
| 6 | Gemini 4 ArgonCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-10-12 | 78.6% | Not reported | Unknown |
| 7 | GPT-5.2Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 76.2% | Not reported | Unknown |
| 8 | GPT-5.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 76.2% | Not reported | Unknown |
| 9 | Gemini 3.7 FlashCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-08-27 | 75.5% | Not reported | Unknown |
| 10 | Grok 4.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-07-15 | 74.5% | Not reported | Unknown |
| 11 | Gemini 2.5 ProCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 74.3% | Not reported | Unknown |
| 12 | Gemini 3.5 FlashCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 73.2% | Not reported | Unknown |
| 13 | Gemini 3.8 FlashCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-09-03 | 73.1% | Not reported | Unknown |
| 14 | Gemini 3.6 FlashCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-07-22 | 72.5% | Not reported | Unknown |
| 15 | Llama 3 - Grounded Language ModelCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-05 | 71.8% | Not reported | Unknown |
| 16 | GPT-5.6 TerraCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-07-15 | 71.2% | Not reported | Unknown |
| 17 | Gemini 3.5 Flash-LiteCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-07-22 | 71.2% | Not reported | Unknown |
| 18 | GPT-5.4Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 70.3% | Not reported | Unknown |
| 19 | Gemini 2.5 FlashCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 70.0% | Not reported | Unknown |
| 20 | Claude Opus 4.8Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 69.7% | Not reported | Unknown |
| 21 | GPT-5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 69.6% | Not reported | Unknown |
| 22 | GPT-5.6 SolCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-07-23 | 68.7% | Not reported | Unknown |
| 23 | Claude Opus 5.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-09-24 | 68.5% | Not reported | Unknown |
| 24 | Gemini 3.1 Flash-Lite PreviewCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 66.5% | Not reported | Unknown |
| 25 | Claude Haiku 5.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2026-10-09 | 65.8% | Not reported | Unknown |
| 26 | Gemini 3.1 Pro PreviewCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 65.7% | Not reported | Unknown |
| 27 | Claude Opus 4.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 62.1% | Not reported | Unknown |
| 28 | Claude Sonnet 4.5 (thinking)Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-10-01 | 61.8% | Not reported | Unknown |
| 29 | Claude Opus 4.7Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 61.7% | Not reported | Unknown |
| 30 | Gemini 3 Flash PreviewCreator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 59.0% | Not reported | Unknown |
| 31 | Claude Sonnet 4.5Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-10-01 | 58.7% | Not reported | Unknown |
| 32 | Gemma 3 12BCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 58.3% | Not reported | Unknown |
| 33 | GPT-5 miniCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-11-14 | 58.3% | Not reported | Unknown |
| 34 | deepseek-v3.1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 57.4% | Not reported | Unknown |
| 35 | Gemma 3 27BCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 56.4% | Not reported | Unknown |
| 36 | Claude Sonnet 4.6Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 56.2% | Not reported | Unknown |
| 37 | Claude Sonnet 4Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 56.1% | Not reported | Unknown |
| 38 | DeepSeek-R1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 56.0% | Not reported | Unknown |
| 39 | Claude Opus 4.6Creator not reportedKaggle benchmark v2; task revision 3; Evaluation date unreported | 54.9% | Not reported | Unknown |
| 40 | Claude Opus 4.1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 54.8% | Not reported | Unknown |
| 41 | Grok 4Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 54.7% | Not reported | Unknown |
| 42 | GPT-5.1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-11-17 | 50.0% | Not reported | Unknown |
| 43 | Mistral Large 2Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 49.7% | Not reported | Unknown |
| 44 | Qwen 3 235B A22B ThinkingCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 48.2% | Not reported | Unknown |
| 45 | GPT-4.1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 45.6% | Not reported | Unknown |
| 46 | Grok 3 MiniCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 45.6% | Not reported | Unknown |
| 47 | Mistral Small 3.1Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-11-13 | 43.3% | Not reported | Unknown |
| 48 | grok-4-fast-reasoningCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-11-14 | 43.1% | Not reported | Unknown |
| 49 | Qwen 3 235B A22B Instruct 2506Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 40.8% | Not reported | Unknown |
| 50 | Gemma 3 4BCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 38.4% | Not reported | Unknown |
| 51 | o3Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 36.2% | Not reported | Unknown |
| 52 | DeepSeek-V3Creator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 33.4% | Not reported | Unknown |
| 53 | gpt-oss-120bCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 30.7% | Not reported | Unknown |
| 54 | o4 miniCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-02 | 29.3% | Not reported | Unknown |
| 55 | gpt-oss-20bCreator not reportedKaggle benchmark v2; task revision 3; Evaluated 2025-09-03 | 23.1% | Not reported | Unknown |
How to read FACTS Grounding v2
FACTS Grounding asks for long-form answers supported by supplied documents of up to 32,000 tokens. Version 2 updates the automated judges. The source describes 1,719 examples across 860 public and 859 held-out cases, and we retain its overall, public, and private scores separately.
Three judges assess factual grounding, with an eligibility check for answers that fail to address the request. The final score averages judge results and the public/private sets. These are document-grounding results under this judge setup; they do not establish open-web factuality or equivalence with version 1.
The complete official download includes models hidden by the source page's default filters. Source names and task versions stay separate from canonical model profiles; these results do not enter overall or category rankings.
Snapshot
These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. The table retains each source configuration and its reported value.
55 source rows are shown for FACTS Grounding v2. The benchmark falls in the Reasoning category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. FACTS Grounding v2 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About FACTS Grounding v2
Year
2025
Tasks
1,719 document-grounded examples: 860 public and 859 held out
Format
Long-form document-grounded answers
FACTS Grounding asks for long-form answers supported by supplied documents of up to 32,000 tokens. Version 2 updates the automated judges. The source describes 1,719 examples across 860 public and 859 held-out cases, and we retain its overall, public, and private scores separately. Three judges assess factual grounding, with an eligibility check for answers that fail to address the request. The final score averages judge results and the public/private sets. These are document-grounding results under this judge setup; they do not establish open-web factuality or equivalence with version 1. The complete official download includes models hidden by the source page's default filters. Source names and task versions stay separate from canonical model profiles; these results do not enter overall or category rankings.
Freshness and provenance
Version
FACTS Grounding v2 2025
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does FACTS Grounding v2 measure?
FACTS Grounding v2 checks whether long-form answers remain supported by supplied documents. Its revised judge ensemble and public/private results remain separate from version 1.
Which source row has the highest reported FACTS Grounding v2 value?
GPT-6 Sol has the highest reported FACTS Grounding v2 value at 85.2%. These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on FACTS Grounding v2?
The October 13, 2026 snapshot contains 55 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.