May 2026

Can AI agents do real-world video post-production work?

We gave the 9 best frontier models 100 expert-authored tasks across the four stages of video post-production. The best agent tops out at 32%. Human experts scored 89%.

100

Tasks

20

Industry experts

7

Frontier models

4

Task families

Why this benchmark exists

Verification is not here for free.

RLVR - reinforcement learning with verifiable rewards - works in math and code because centuries of humanistic work built the verifiers; the bill was paid before we got there. Creative work hasn't paid that bill. AgenticVBench is what paying it looks like in film.

It also measures the sim2real gap: the distance between how agents score on tidy lab benchmarks and how they hold up on real post-production work. Here that gap is stark - the best frontier agent scores 32%, human experts 89%.

Read the full essay →

Leaderboard preview

Top 5 model × harness combinations.

View full leaderboard →
RankAgentAvg
·Human expertsreference88.5%
1Claude Fable 5· Claude Code32.4%± 5.1
2GPT-5.5· Codex31.0%± 4.0
3GPT-5.5· OpenCode27.4%± 3.5
4Gemini 3.1 Pro· OpenCode23.8%± 3.7
5MiniMax-M3· OpenCode22.7%± 3.2

The harness finding

The harness matters as much as the model.

Holding the model fixed and varying the harness shifts GPT-5.5's Assembly score by 20 percentage points, comparable to the gap between adjacent models on the leaderboard.

Most benchmarks today are still model-based. The data here says that's wrong. Agent performance is determined by both the model and the scaffolding around it. Reporting only the model misses the larger story.

Agent = model × harness.

GPT-5.5 on Assembly · score by harness

Codex
38%
OpenCode
37%
OpenClaw
18%

Same model. 20-point swing.

Cite this work

Citation

If you find AgenticVBench useful in your research, please consider citing the paper.

BibTeX
@article{cao2026agenticvbench,
  title={AgenticVBench: Can AI Agents Complete Real-World Post-Production Tasks?},
  author={Cao, Zongheng and Zheng, Yi and Song, Rui and Hu, Xinyu},
  journal={arXiv preprint arXiv:2605.27705},
  year={2026}
}
Read the paper on arXiv →