MMTB: leaderboard
MultiMedia-TerminalBench (Heo et al., 2026): 105 terminal-agent tasks based on paid practitioner workflows, in which the agent must inspect audio, video and image files in a persistent workspace and write a verifiable output. The score is the percentage of tasks whose verifier passes in full, one run per task with a 10-minute budget; higher is better. Rows are agent harness and model configurations, including harnesses with native audio and video access.
Metric: Binary task success rate (%; share of the 105 multimedia-file terminal tasks whose verifier passes in full; one run per task with a 10-minute agent budget; harness and model configuration as labelled). Source: arxiv.org. Saturation forecast: Around August 2028. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 16.2 | #158 |
| 2 | Gemini 3.1 Pro (Preview) | 12.4 | #91 |
| 3 | GPT-5.2 | 11.4 | #172 |
| 4 | Qwen 3.5 122B A10B | 10.5 | #329 |
| 5 | Gemini 2.5 Flash | 6.7 | #452 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mmtb · How It Works · Data refreshed daily, snapshot 2026-11-12.