SDABench: leaderboard

Metric: Accuracy (%) of open-ended answers on the held-out SDA-Synth test split (2,400 instances, 400 per task type across five scientific domains, macro-averaged over the six task types; zero-shot, JSON reasoning and answer fields, the answer scored by relative-error numeric, exact or structured matching with a GPT-4o fallback judge); higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.

Top models

#ModelScore
1GPT-5.458.33
2Gemini 3.1 Pro (Preview)56.88
3Claude Sonnet 4.655.79
4GLM-551.67
5Qwen 3.5 397B A17B51.58
6DeepSeek V3.251.17
7Kimi K2.550.29
8DeepSeek R142.13
9Qwen 3 235B A22B 2507 Instruct40.46
10GPT-5.4 Mini37.67
11Llama 3.3 70B Instruct6.96
12Llama 3.1 8B Instruct5.04

Interactive version: theaggregate.ai/benchmark?slug=sdabench · How It Works · Data refreshed daily, snapshot 2026-09-29.