HellaSwag: leaderboard, scores and methodology
Commonsense sentence completion with adversarial wrong answers.
Top 10
Bars are scaled to the highest observed score. The leader sets the pace; the spread below it shows how much separation exists in this benchmark.
Leaderboard
| # | Model | Provider | Context | Input / 1M | Avg score |
|---|---|---|---|---|---|
| 1 | GPT-5.2 Max | OpenAI | 1000000 | $21.39 | 89.2 |
| 2 | GPT-5.2 Turbo | OpenAI | 1000000 | $20.73 | 87.4 |
| 3 | GPT-5.2 Pro | OpenAI | 2000000 | $21.32 | 89.6 |
| 4 | Claude Opus 4.5 Pro | Anthropic | 200000 | $20.98 | 89.1 |
| 5 | Claude Sonnet 4.5 Reasoning | Anthropic | 400000 | $19.48 | 87.7 |
| 6 | Grok 4 Reasoning | xAI | 2000000 | $19.67 | 87.8 |
| 7 | GPT-5.2 Vision | OpenAI | 1000000 | $19.03 | 87.4 |
| 8 | Gemini 3 Pro Turbo | Google DeepMind | 400000 | $19.62 | 86.4 |
| 9 | GPT-5.1 Reasoning | OpenAI | 400000 | $20.27 | 88.3 |
| 10 | o4 Pro | OpenAI | 1000000 | $19.37 | 87.3 |
| 11 | Grok 4 Max | xAI | 2000000 | $19.70 | 86.6 |
| 12 | Claude Sonnet 4.5 Turbo | Anthropic | 1000000 | $18.37 | 85.9 |
| 13 | Grok 4 | xAI | 200000 | $17.90 | 86.5 |
| 14 | Claude Opus 4.5 Reasoning | Anthropic | 2000000 | $20.91 | 88.2 |
| 15 | GPT-5.1 Pro | OpenAI | 200000 | $19.77 | 87.8 |
| 16 | GPT-4.1 Pro | OpenAI | 400000 | $17.56 | 85.2 |
| 17 | Gemini 3 Pro Pro | Google DeepMind | 200000 | $19.80 | 87.3 |
| 18 | Gemini 3 Pro Reasoning | Google DeepMind | 200000 | $19.69 | 87.5 |
| 19 | o3 Max | OpenAI | 2000000 | $17.82 | 84.8 |
| 20 | Claude Sonnet 4.5 Max | Anthropic | 1000000 | $19.90 | 87.2 |
| 21 | GPT-5.1 Max | OpenAI | 2000000 | $20.50 | 87.9 |
| 22 | Claude Opus 4.5 Max | Anthropic | 1000000 | $20.90 | 87.8 |
| 23 | Claude Sonnet 4.5 | Anthropic | 400000 | $17.84 | 86.0 |
| 24 | Gemini 2.5 Pro Max | Google DeepMind | 200000 | $16.94 | 84.8 |
| 25 | Qwen3-Max | Alibaba Cloud | 400000 | $17.37 | 84.4 |
| 26 | Claude Opus 4.5 Turbo | Anthropic | 2000000 | $19.57 | 85.7 |
| 27 | Grok 4 Pro | xAI | 1000000 | $19.00 | 85.4 |
| 28 | o3 | OpenAI | 400000 | $16.80 | 84.5 |
| 29 | Claude Sonnet 4.5 Pro | Anthropic | 1000000 | $18.73 | 86.7 |
| 30 | GPT-5.2 | OpenAI | 2000000 | $19.71 | 88.0 |
| 31 | Gemini 3 Pro Max | Google DeepMind | 2000000 | $20.57 | 87.9 |
| 32 | GPT-5.2 Reasoning | OpenAI | 1000000 | $20.98 | 87.5 |
| 33 | Grok 4 Turbo | xAI | 1000000 | $19.21 | 86.7 |
| 34 | o3 Pro | OpenAI | 400000 | $17.89 | 85.8 |
| 35 | Claude Opus 4.5 | Anthropic | 400000 | $19.42 | 87.7 |
| 36 | Qwen3-Max Pro | Alibaba Cloud | 400000 | $18.87 | 86.2 |
| 37 | o4 Reasoning | OpenAI | 1000000 | $19.69 | 87.5 |
| 38 | Kimi K2 | Moonshot AI | 128000 | $5.85 | 83.0 |
| 39 | GPT-5.1 | OpenAI | 2000000 | $19.01 | 86.0 |
| 40 | Gemini 3 Flash Reasoning | Google DeepMind | 128000 | $17.05 | 83.7 |
| 41 | GPT-5.1 Vision | OpenAI | 400000 | $17.76 | 84.8 |
| 42 | Gemini 2.5 Pro Reasoning | Google DeepMind | 128000 | $16.81 | 83.8 |
| 43 | Grok 3 Max | xAI | 2000000 | $17.66 | 85.9 |
| 44 | o3 Turbo | OpenAI | 200000 | $16.63 | 83.4 |
| 45 | o4 Max | OpenAI | 2000000 | $19.51 | 86.2 |
| 46 | Kimi K2 Chat | Moonshot AI | 400000 | $5.74 | 82.1 |
| 47 | Gemini 3 Flash Vision | Google DeepMind | 400000 | $15.06 | 81.9 |
| 48 | Qwen3-Max Reasoning | Alibaba Cloud | 2000000 | $18.12 | 85.9 |
| 49 | Claude Opus 4.5 Vision | Anthropic | 200000 | $17.70 | 85.9 |
| 50 | Grok 3 | xAI | 128000 | $15.84 | 82.8 |
Methodology
HellaSwag is included as a commonsense benchmark with a typical score range of 0 to 100. The values in this index are deterministic modelled estimates generated for comparison at scale. They are not official leaderboard results and should not be quoted as vendor benchmarks.
Each model receives a score derived from its capability tier, release generation and benchmark difficulty. Scores are normalised to a 0–100 scale so that models from different providers can be placed side by side. The average shown is the arithmetic mean across all 600 scored models.
Read the leaderboard as a relative map. A five-point gap is meaningful at the top and less meaningful in the middle of the distribution. Pair the score with price and context data before choosing a model.
How to use this leaderboard
- Use the top 10 to shortlist candidates, not to declare a winner.
- Check whether the leader is open weights or proprietary before assuming you can deploy it.
- Compare the score against input and output price on each model page.
- Look at the model's scores on related benchmarks before trusting a single number.
FAQ
What is HellaSwag?
Commonsense sentence completion with adversarial wrong answers.
Who leads HellaSwag?
GPT-5.2 Max from OpenAI leads this index with a score of 96.1, followed by GPT-5.2 Turbo at 95.5.
What is a good HellaSwag score?
The top quartile starts around 87.8. The index average is 81.4, so a score above 87.8 is a strong result relative to this dataset.
How are HellaSwag scores calculated?
The values are modelled estimates on a 0–100 scale, generated from capability tier, release generation and benchmark difficulty. They are not official leaderboard results.
Which models are included?
All 600 models in the index are scored on HellaSwag, across 40 providers and every model category.