InteractBench - Hard - pass@1: leaderboard
Metric: pass@1 (%; n=10 samples per task). Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 36.7 |
| 2 | GPT-5.2 (Thinking) | 34 |
| 3 | DeepSeek V3.2 (Thinking) | 10.4 |
| 4 | Gemini 2.5 Pro | 9.2 |
| 5 | GPT-OSS-120B | 7.8 |
| 6 | DeepSeek V3.2 (Non-reasoning) | 2.4 |
| 7 | GPT-OSS-20B | 2.1 |
| 8 | Claude Opus 4.5 (Thinking) | 2.1 |
| 9 | Qwen 3 30B A3B 2507 (Thinking) | 1.2 |
| 10 | Gemini 2.5 Flash | 0.9 |
| 11 | Qwen 3 32B (Thinking) | 0.9 |
| 12 | GPT-4.1 | 0.3 |
| 13 | DeepSeek R1 Distill Llama 70B | 0 |
| 14 | Qwen 3 14B (Non-reasoning) | 0 |
| 15 | Qwen 3 32B (Non-reasoning) | 0 |
Interactive version: theaggregate.ai/benchmark?slug=interactbench-hard-pass-1 · How It Works · Data refreshed daily, snapshot 2026-09-19.