ESG Benchmark - Level 2: leaderboard
Level 2 of the ESG Benchmark (Zhao et al., 2026): 114 compositional ESG tasks needing three to five reasoning steps, such as carbon and financial calculations, case analysis and cross-document comparison, with PDF, CSV, MP3 and MP4 attachments. Accuracy (%) is computed from the paper's printed correct-answer counts out of 114; higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around October 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 71.05 | #159 |
| 2 | Gemini 3 Pro | 66.67 | #135 |
| 3 | Gemini 2.5 Pro | 65.79 | #262 |
| 4 | GPT-5 | 65.79 | #163 |
| 5 | DeepSeek R1 | 65.79 | #467 |
| 6 | O3 | 65.79 | #224 |
| 7 | Claude Opus 4.5 | 64.91 | #152 |
| 8 | Gemini 2.5 Flash | 64.04 | #452 |
| 9 | GPT-4.1 | 64.04 | #492 |
| 10 | GPT-4o | 62.28 | #741 |
| 11 | DeepSeek V3 | 62.28 | #705 |
| 12 | Claude Haiku 4.5 | 60.53 | #568 |
| 13 | GPT-5.2 | 58.77 | #172 |
| 14 | GPT-5.1 | 58.77 | #225 |
| 15 | Claude Sonnet 4.5 | 57.89 | #244 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=esg-benchmark-level-2 · How It Works · Data refreshed daily, snapshot 2026-11-12.