EsoBench: leaderboard
Can models learn an unknown programming language from scratch? 30 tasks across 6 difficulty levels with only 1 example and 50 experimental attempts. No model has fully solved it - top score is 29.4/100.
Metric: Score. Source: caseys-evals.com. Saturation forecast: Around 2030. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 (Thinking) | 29.4 | #260 (Qwen 3 235B A22B 2507) |
| 2 | GPT-5 | 28.13 | #98 |
| 3 | GPT-5.2 (2025-12-11) (High) | 26.87 | |
| 4 | O4 Mini (2025-04-16) (High) | 26.8 | #179 (O4 Mini (2025-04-16)) |
| 5 | GPT-5 Mini (2025-08-07) | 23.73 | #172 |
| 6 | O4 Mini (2025-04-16) | 23.07 | #179 |
| 7 | Grok 4 | 21.4 | #181 |
| 8 | Claude Opus 4.5 (20251101) | 20.93 | #74 |
| 9 | GPT-5.1 (2025-11-13) (High) | 19.33 | #101 (GPT-5.1 (2025-11-13)) |
| 10 | Claude Opus 4.1 (20250805) | 19.27 | #132 |
| 11 | Gemini 3 Pro (Preview) | 17.73 | #66 |
| 12 | Claude Sonnet 4.6 | 17.4 | #90 |
| 13 | Grok 3 | 16.8 | #304 |
| 14 | Claude Opus 4 (20250514) | 16.8 | #154 |
| 15 | O3 Mini (2025-01-31) | 16.67 | #240 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=esobench · How It Works · Data refreshed daily, snapshot 2026-10-20.