GAIA2 - Execution: leaderboard
GAIA2 (Meta Agents Research Environments) capability slice measuring an agent's ability to carry out multi-step tasks by executing the correct actions and tool calls.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Around January 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 69.2 | #98 (GPT-5) |
| 2 | Claude Sonnet 4 (Thinking) | 62.1 | #199 (Claude Sonnet 4) |
| 3 | Claude Sonnet 4 | 57.9 | #199 |
| 4 | GPT-5 (Low) | 52.7 | #98 (GPT-5) |
| 5 | DeepSeek V3.1 Terminus | 43.1 | #220 |
| 6 | DeepSeek V3.1 | 39.8 | #264 |
| 7 | Gemini 2.5 Pro | 39.2 | #152 |
| 8 | Kimi K2 | 34.2 | #243 |
| 9 | GPT-5 (Minimal) | 31.9 | #98 (GPT-5) |
| 10 | Qwen 3 235B A22B (Thinking) | 28.1 | #313 (Qwen 3 235B A22B) |
| 11 | Qwen 3 235B A22B | 22.7 | #313 |
| 12 | GPT-OSS-120B (High) | 17.9 | #341 (GPT-OSS-120B) |
| 13 | Llama 4 Maverick | 13.8 | #464 |
| 14 | Grok 4 | 8.8 | #181 |
| 15 | GPT-4o | 8.3 | #343 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gaia2-execution · How It Works · Data refreshed daily, snapshot 2026-10-20.