InteractBench - Overall - pass@1: leaderboard
InteractBench (arXiv 2608.29632v1, Table 2): interactive competitive-programming tasks where the submitted program must query a judge program under a protocol and query budget. pass@1 estimated from 10 samples per task at temperature 0.8 on all problems, in percent; higher is better. Thinking and non-thinking runs of one checkpoint are separate rows.
Metric: pass@1 (%; n=10 samples per task). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 55.4 | #71 |
| 2 | GPT-5.2 (Thinking) | 54.2 | #115 (GPT-5.2) |
| 3 | DeepSeek V3.2 (Thinking) | 33.2 | #210 (DeepSeek V3.2) |
| 4 | Gemini 2.5 Pro | 30.5 | #159 |
| 5 | GPT-OSS-120B | 25.4 | #346 |
| 6 | Gemini 2.5 Flash | 16.2 | #255 |
| 7 | Claude Opus 4.5 (Thinking) | 15.9 | #87 (Claude Opus 4.5) |
| 8 | DeepSeek V3.2 (Non-reasoning) | 14.8 | #210 (DeepSeek V3.2) |
| 9 | GPT-OSS-20B | 13.4 | #523 |
| 10 | Qwen 3 30B A3B 2507 (Thinking) | 10.5 | #379 |
| 11 | Qwen 3 32B (Thinking) | 9.6 | #449 (Qwen 3 32B) |
| 12 | GPT-4.1 | 6.2 | #258 |
| 13 | Qwen 3 14B (Reasoning) | 5.8 | #542 (Qwen 3 14B) |
| 14 | DeepSeek R1 Distill Llama 70B | 4.5 | #614 |
| 15 | Qwen 3 32B (Non-reasoning) | 1.7 | #449 (Qwen 3 32B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=interactbench-overall-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-05.