OI-Bench - Standard Accuracy: leaderboard
OI-Bench (Liou et al., 2026) control condition: zero-shot four-option accuracy (%) on 3,000 questions from the tinyBenchmarks subsets of MMLU, LogiQA 2.0 and HellaSwag, answered with greedy decoding in a fixed answer format. Higher is better. No injected option is present, so this measures knowledge, logic and commonsense rather than robustness.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 89.43 | #163 |
| 2 | Gemini 2.5 Pro | 87.66 | #262 |
| 3 | DeepSeek R1 | 86.92 | #467 |
| 4 | Qwen 3 235B A22B | 84.26 | #622 |
| 5 | DeepSeek V3.2 | 83.94 | #403 |
| 6 | Llama 4 Maverick | 82.07 | #1105 |
| 7 | Claude Haiku 4.5 | 79.84 | #568 |
| 8 | Grok 4.1 | 78.08 | #473 |
| 9 | Llama 4 Scout | 78.05 | #1787 |
| 10 | GPT-5 Mini | 76.08 | #349 |
| 11 | Gemini 2.5 Flash Lite | 69.57 | #958 |
| 12 | Qwen 3 8B | 69.3 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=oi-bench-standard-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.