Leaderboard
Self-reported vs Reproduced
7,285 sourced scores across 1,916 models and 7,569 benchmarks. Nothing here is estimated or interpolated; a missing cell means no sourced number. Community reproductions appear on each benchmark page and in recent runs.
Composite
A weighted average over the ten capability axes. Only shown when coverage is sufficient; otherwise N/A with the coverage reason.
| # | Model | Composite | Coverage & evidence | Axes |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 88.7 | 59/668 benchmarks · 8/10 axes 95% high/med confidence 0% self-reported · 100% independent |
|
| 2 | Claude Sonnet 5.5Anthropic | 82.6 | 8/668 benchmarks · 3/10 axes 82% high/med confidence 0% self-reported · 100% independent |
|
| 3 | Claude Fable 5Anthropic | 76.0 | 13/668 benchmarks · 6/10 axes 89% high/med confidence 0% self-reported · 100% independent |
|
| 4 | Gemini 3 Flash PreviewGoogle | 75.5 | 28/668 benchmarks · 6/10 axes 97% high/med confidence 0% self-reported · 100% independent |
|
| 5 | GPT-5.4OpenAI | 75.4 | 47/668 benchmarks · 8/10 axes 81% high/med confidence 10% self-reported · 90% independent |
|
| 6 | Claude Opus 4.7Anthropic | 75.3 | 19/668 benchmarks · 6/10 axes 83% high/med confidence 0% self-reported · 100% independent |
|
| 7 | Claude Opus 4.6Anthropic | 73.3 | 34/668 benchmarks · 8/10 axes 65% high/med confidence 31% self-reported · 69% independent |
|
| 8 | GPT-5.6 SolOpenAI | 72.5 | 106/668 benchmarks · 10/10 axes 98% high/med confidence 0% self-reported · 100% independent |
|
| 9 | GPT-6.1 SolOpenAI | 72.4 | 10/668 benchmarks · 6/10 axes 85% high/med confidence 0% self-reported · 100% independent |
|
| 10 | Gemini 3.6 FlashGoogle | 72.1 | 15/668 benchmarks · 6/10 axes 89% high/med confidence 0% self-reported · 100% independent |
|
| 11 | GPT-5.5OpenAI | 71.9 | 51/668 benchmarks · 9/10 axes 93% high/med confidence 0% self-reported · 100% independent |
|
| 12 | Claude Fable 5.1Anthropic | 71.7 | 20/668 benchmarks · 7/10 axes 87% high/med confidence 0% self-reported · 100% independent |
|
| 13 | Claude Opus 5Anthropic | 71.5 | 59/668 benchmarks · 8/10 axes 97% high/med confidence 0% self-reported · 100% independent |
|
| 14 | Claude Opus 4.5 (2025-11-01)Anthropic | 69.9 | 12/668 benchmarks · 5/10 axes 84% high/med confidence 8% self-reported · 92% independent |
|
| 15 | Gemini 3.7 FlashGoogle | 69.0 | 22/668 benchmarks · 6/10 axes 93% high/med confidence 0% self-reported · 100% independent |
|
| 16 | Gemini 3.8 FlashGoogle | 68.7 | 51/668 benchmarks · 10/10 axes 95% high/med confidence 0% self-reported · 100% independent |
|
| 17 | Claude Opus 5.5Anthropic | 68.4 | 32/668 benchmarks · 8/10 axes 94% high/med confidence 0% self-reported · 100% independent |
|
| 18 | Claude Sonnet 4.6Anthropic | 66.1 | 39/668 benchmarks · 8/10 axes 66% high/med confidence 28% self-reported · 72% independent |
|
| 19 | Qwen2.5-VL-7BAlibaba | 65.9 | 25/668 benchmarks · 3/10 axes 100% high/med confidence 0% self-reported · 100% independent |
|
| 20 | Claude Opus 4.8Anthropic | 65.8 | 45/668 benchmarks · 9/10 axes 91% high/med confidence 6% self-reported · 94% independent |
|
| 21 | GPT-5.6-terraOpenAI | 64.8 | 39/668 benchmarks · 9/10 axes 95% high/med confidence 0% self-reported · 100% independent |
|
| 22 | Qwen3-VL-8B-InstructAlibaba | 64.5 | 41/668 benchmarks · 5/10 axes 100% high/med confidence 0% self-reported · 100% independent |
|
| 23 | GPT-6 SolOpenAI | 63.9 | 11/668 benchmarks · 6/10 axes 86% high/med confidence 0% self-reported · 100% independent |
|
| 24 | Gemini 3 ProGoogle | 63.4 | 20/668 benchmarks · 8/10 axes 54% high/med confidence 30% self-reported · 70% independent |
|
| 25 | Gemini-3.5-flashGoogle | 63.1 | 29/668 benchmarks · 7/10 axes 94% high/med confidence 0% self-reported · 100% independent |
|
| 26 | Grok-4.7xAI | 63.1 | 11/668 benchmarks · 6/10 axes 86% high/med confidence 0% self-reported · 100% independent |
|
| 27 | Gemini 3.1 ProGoogle | 60.5 | 68/668 benchmarks · 8/10 axes 77% high/med confidence 16% self-reported · 84% independent |
|
| 28 | Qwen Max (2025-01-25)Alibaba | 59.9 | 10/668 benchmarks · 6/10 axes 100% high/med confidence 0% self-reported · 100% independent |
|
| 29 | Qwen3.6 27BAlibaba | 59.3 | 15/668 benchmarks · 9/10 axes 94% high/med confidence 0% self-reported · 100% independent |
|
| 30 | Qwen 3.5Alibaba | 59.2 | 14/668 benchmarks · 5/10 axes 14% high/med confidence 86% self-reported · 14% independent |
|
| 31 | Claude Sonnet 4.5 (2025-09-29)Anthropic | 59.1 | 15/668 benchmarks · 7/10 axes 84% high/med confidence 3% self-reported · 97% independent |
|
| 32 | Grok-4.6xAI | 57.8 | 16/668 benchmarks · 7/10 axes 90% high/med confidence 0% self-reported · 100% independent |
|
| 33 | Gemini 3.5 Flash LiteGoogle | 57.4 | 8/668 benchmarks · 6/10 axes 82% high/med confidence 0% self-reported · 100% independent |
|
| 34 | Kimi K3Moonshot AI | 57.3 | 55/668 benchmarks · 8/10 axes 97% high/med confidence 0% self-reported · 100% independent |
|
| 35 | Qwen3.8 MaxAlibaba | 57.3 | 39/668 benchmarks · 8/10 axes 98% high/med confidence 0% self-reported · 100% independent |
|
| 36 | DeepSeek ChatDeepSeek | 56.8 | 13/668 benchmarks · 7/10 axes 100% high/med confidence 0% self-reported · 100% independent |
|
| 37 | GPT-5OpenAI | 55.7 | 18/668 benchmarks · 8/10 axes 85% high/med confidence 0% self-reported · 100% independent |
|
| 38 | Qwen3.7 MaxAlibaba | 55.3 | 13/668 benchmarks · 6/10 axes 93% high/med confidence 0% self-reported · 100% independent |
|
| 39 | Qwen3.7 PlusAlibaba | 54.7 | 11/668 benchmarks · 5/10 axes 92% high/med confidence 0% self-reported · 100% independent |
|
| 40 | Kimi K2.5Moonshot AI | 54.4 | 31/668 benchmarks · 8/10 axes 64% high/med confidence 23% self-reported · 77% independent |
|
| 41 | GPT-5.1OpenAI | 54.4 | 12/668 benchmarks · 6/10 axes 76% high/med confidence 0% self-reported · 100% independent |
|
| 42 | google/gemma-4-31B-itGoogle | 54.3 | 24/668 benchmarks · 7/10 axes 96% high/med confidence 0% self-reported · 100% independent |
|
| 43 | GPT-5.2-chatOpenAI | 54.2 | 21/668 benchmarks · 7/10 axes 88% high/med confidence 0% self-reported · 100% independent |
|
| 44 | Qwen3.8 27BAlibaba | 54.0 | 30/668 benchmarks · 8/10 axes 94% high/med confidence 0% self-reported · 100% independent |
|
| 45 | qwen3.5-27BAlibaba | 52.2 | 22/668 benchmarks · 7/10 axes 100% high/med confidence 0% self-reported · 100% independent |
|
| 46 | GPT-5.4-miniOpenAI | 52.1 | 20/668 benchmarks · 8/10 axes 92% high/med confidence 0% self-reported · 100% independent |
|
| 47 | DeepSeek V4.1 FlashDeepSeek | 51.8 | 40/668 benchmarks · 8/10 axes 96% high/med confidence 0% self-reported · 100% independent |
|
| 48 | Claude 3.5 SonnetAnthropic | 51.8 | 12/668 benchmarks · 6/10 axes 96% high/med confidence 0% self-reported · 100% independent |
|
| 49 | DeepSeek V3.2DeepSeek | 51.8 | 24/668 benchmarks · 5/10 axes 66% high/med confidence 24% self-reported · 76% independent |
|
| 50 | o3OpenAI | 51.4 | 13/668 benchmarks · 7/10 axes 76% high/med confidence 0% self-reported · 100% independent |
|
How the composite is computed
0. Only benchmarks that are still being measured count: a benchmark needs a dated score within 730 days of the newest score in the dataset, and at least 3 models scored on it. Stale benchmarks measured only on older models would otherwise inflate those models.
1. On each benchmark, a model's best score is turned into a percentile rank among all models scored there (100 = best, 0 = worst; lower-is-better metrics are inverted). Ranks, not raw numbers, so an easy benchmark does not outweigh a hard one.
2. For each model, the best score per benchmark is used. In Any harness mode this is across all harnesses; in Grouped mode each (model, harness) pair is scored separately.
3. Each capability axis is the weighted mean of its benchmarks (a benchmark counts under its primary axis). Saturated benchmarks are down-weighted to 0.5x because they discriminate less.
4. The composite is the weighted mean over covered axes with weights Knowledge & Reasoning 1, Math 1, Coding 1.5, Agentic / Tool Use 1.5, Multimodal 0.75, Long Context 0.75, Instruction Following 0.75, Safety & Alignment 0.5, Abstract Reasoning 0.75, Multilingual 0.5. A model needs at least 5 benchmarks across 3 axes, otherwise the composite is N/A.
Source types counted: Self-reported, Independent, Reproduced. The same algorithm ships in the shared package (@ltb/shared, computeCompositeScores) and in the CLI, so numbers match everywhere.
Model × benchmark
Columns are benchmarks grouped by primary axis; cells show the normalized score (0–100) of the best included source. Hover or focus a cell for the raw value and source link.