VisEditBench - Overall: leaderboard
VisEditBench (arXiv 2608.10408v1, Table 2): zero-shot vision-language models revise Vega-Lite or Matplotlib code from a chart image, feedback or reference chart and an instruction, over 1,395 human-annotated tasks. A pass needs executable code, task accuracy of at least 4.5 of 5, readability and visual quality of 4.0 and visual similarity of 90 under a GPT-4o evaluator (Gemini 2.5 Pro for GPT-4o outputs). Pass rate in percent; higher is better.
Metric: Pass rate (%; all 1,395 examples). Source: arxiv.org. Saturation forecast: Around February 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 74.46 | #94 |
| 2 | GPT-5 | 65.68 | #105 |
| 3 | Claude Sonnet 4.5 | 62.73 | #148 |
| 4 | GPT-5 Mini | 60 | #193 |
| 5 | GPT-4o | 55.75 | #357 |
| 6 | Gemma 3 27B | 42.95 | #606 |
| 7 | Pixtral-12B | 35.32 | #805 |
| 8 | Gemma 3 12B | 34.96 | #674 |
| 9 | Gemma 3 4B | 21.29 | #1088 |
| 10 | Gemini 3 Flash | 17.06 | #103 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=viseditbench-overall · How It Works · Data refreshed daily, snapshot 2026-11-05.