ChartAct: leaderboard

ChartAct (2026) tests dynamic chart understanding: a multimodal agent sees only screenshots and must hover, click, zoom or drag real interactive web charts to reveal the evidence for each of 1,440 questions, once on a clean page and once inside a dashboard. Success rate in percent averaged over the two environments; three costly models are scored on a 300-case representative subset; higher is better.

Metric: Success rate (%): mean of the Dynamic Chart and Dashboard Chart success rates; the model acts in an OSWorld-style desktop on screenshots only (pyautogui actions, at most 15 steps, temperature 0.3) to hover, click, zoom or drag 673 real interactive web charts and answers 1,440 human-reviewed questions, an LLM judge with majority voting deciding correctness; Claude Opus 4.7, GPT-5.5 and Gemini 3.1 Pro are scored on a representative 300-case subset per environment; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.784.5#85
2Kimi K2.5Non-reasoning66#266
3GPT-5.558.3#54
4Doubao-Seed-2-0-Pro-260215High47.5
5Gemini 3.1 Pro (Preview)Low46#99
6Qwen 3.5 122B A10B41.9#346
7Qwen 3.6 PlusThinking39.2#238
8Qwen 3.5 35B A3B33.1#492
9GLM-4.6V30.8#710
10Gemma 4 31B (IT)23.1#350
11Qwen 3 VL PlusThinking22.5#512

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=chartact · How It Works · Data refreshed daily, snapshot 2026-11-21.