SaaS-Bench - Multimodal: leaderboard

SaaS-Bench (2026) evaluates computer-use agents on long professional workflows across 23 self-hosted SaaS systems, scoring weighted verification checkpoints on the final application state. This board is the checkpoint score in percent on the tasks that use images, documents or media files; higher is better.

Metric: Checkpoint score (%) on the 32 multimodal tasks (images, documents and media files as inputs), weighted-checkpoint score over the final application states, browser-use agent loop over live self-hosted SaaS apps (DOM and screenshot observations, restricted browser actions, no JavaScript or backend access), same step budget, timeout and failure cap for every model; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.648.7#114
2GPT-5.5High47.6#56
3Claude Opus 4.746.3#86
4GPT-5.4High46.1#138
5Qwen 3.6 Plus45.5#245
6Seed 2.0 Pro44.2#226
7Kimi K2.643.5#197
8Gemini 3.1 Pro (Preview)42.4#102
9Kimi K2.537#272
10Gemini 3.5 FlashHigh36.5#110
11Claude Sonnet 4.633.9#179

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=saas-bench-multimodal · How It Works · Data refreshed daily, snapshot 2026-11-22.