AIDABench - Hard: leaderboard
Metric: Score (%) on the 127 hard tasks (13 or more expert operation steps), best of three runs per task (pass@3), AIDABench's 603 end-to-end data-analytics tasks over real business files (mostly spreadsheets, up to 14 files per task); the model works in a Python sandbox through one execute_code tool (at most 20 calls) with an auxiliary spreadsheet summary, provider-recommended decoding; QA answers are judged pass/fail by QwQ-32B against the reference, charts get a 0-1 rubric score from a Gemini-3-Pro-Preview judge, spreadsheets pass/fail from a Claude-Sonnet-4.5 file evaluator; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 50.54 | #138 |
| 2 | Gemini 3 Pro (Preview) | 46.65 | #64 |
| 3 | DeepSeek V3.2 (Non-reasoning) | 39.51 | #198 (DeepSeek V3.2) |
| 4 | DeepSeek V3.2 (Thinking) | 38.1 | #198 (DeepSeek V3.2) |
| 5 | Qwen 3 235B A22B 2507 Instruct | 37.92 | #291 |
| 6 | Qwen 3 Max (2026-01-23) | 36.6 | #151 |
| 7 | Kimi K2 (Thinking) | 34.69 | #236 (Kimi K2) |
| 8 | Qwen 3 30B A3B 2507 (Thinking) | 15.31 | #366 (Qwen 3 30B A3B 2507) |
| 9 | Qwen 3 30B A3B 2507 Instruct | 12.39 | #464 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=aidabench-hard · How It Works · Data refreshed daily, snapshot 2026-10-11.