Cinecourier
Analysis

VIVID harness lifts Codex to 58.6% on AgenticVBench, authors report

A paper's visual harness for coding agents scores 58.6% on a 100-task video editing benchmark; we review ablations, MIT licence and API costs.

According to the VIVID authors, on AgenticVBench, a 100-task benchmark covering Assembly, Repair, Sequencing and Repurpose, Codex with VIVID averages 58.6%, which is 20.2 points above native Codex; Claude Code and OpenCode with VIVID reach 55.1% and 37.1% (arXiv). VIVID is a visual harness that lets several native coding agents handle video post-production, with no extra model training, according to the paper. For practitioners the useful follow-up is what such a gain costs in tokens, and the paper’s own cost table gives a partial answer.

What the harness is

VIVID does not train a new model. It wraps coding agents that already exist, which the paper calls native agents, and adds tooling for working with video. The repository is licensed under MIT (GitHub). For a team that already runs a coding agent, a permissive licence and no training requirement lower the cost of trying the approach.

Same-model comparison and ablations

In an additional same-model comparison using GPT-5.6-sol, the paper says Codex + VIVID scores 58.6, against 38.4 for native Codex and 33.7 for Crayotter (arXiv). Holding the model fixed matters, because it isolates the effect of the harness from the effect of the underlying model.

The authors also compare full VIVID with three variants, all using GPT-5.6-sol with Codex. Taking out on-demand audiovisual inspection lowers the score to 42.6, down 16.0 percentage points. Without persistent evidence memory and retrieval, the score is 47.8. In our reading, the ablation points to the ability to look at the video and audio on demand as the component that matters most in this setup.

The authors ran every task three times, independently, for each model–harness pairing, with a fixed tool schema.

Token use and cost per run

Published resultsTable 8 of the paper: Token usage and estimated cost by harness and task groupRows of Table 8 of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column"; an empty name cell repeats the name above it. These are the authors' own figures, not an independent test.Data read November 17, 2026Source [2]How this format worksDownload the data behind it
Harness Task group Input (M) Output (M) Cost ($) Cost/run ($) Source page
Codex Non-Repurpose 364.12 2.75 253.56 1.32 https://arxiv.org/html/2611.20688
+ GPT-5.6-sol Repurpose 344.25 1.92 217.47 2.01 https://arxiv.org/html/2611.20688
Claude Code Non-Repurpose 274.21 3.64 607.32 3.16 https://arxiv.org/html/2611.20688
+ Claude Fable 5 Repurpose 248.17 3.57 570.87 5.29 https://arxiv.org/html/2611.20688
OpenCode Non-Repurpose 335.86 6.04 600.72 3.13 https://arxiv.org/html/2611.20688
+ GPT-5.5 Repurpose 114.11 2.20 228.92 2.12 https://arxiv.org/html/2611.20688

The authors accessed model APIs through OpenRouter and report token usage and estimated costs for three model–harness configurations. Repurpose is reported on its own because its evaluators are Gemini 3.1 Pro and Claude Opus 4.7. Per-run cost is the total cost divided by the number of runs, and costs are in US dollars.

For Codex with GPT-5.6-sol, non-Repurpose tasks cost 253.56 dollars in total and 1.32 dollars per run. For Claude Code with Claude Fable 5, Repurpose cost 5.29 dollars per run. For OpenCode with GPT-5.5, non-Repurpose tasks cost 3.13 dollars per run. Reading the table, Codex is the cheapest configuration per run on non-Repurpose tasks at 1.32 dollars, against 3.16 for Claude Code and 3.13 for OpenCode. Input volume does not track cost in Table 8: on non-Repurpose tasks Codex with GPT-5.6-sol used 364.12 million input tokens, more than OpenCode with GPT-5.5 at 335.86 million, yet cost 253.56 dollars against 600.72 dollars for OpenCode. The Claude Code Repurpose row is the costliest per run at 5.29 dollars, while the Codex and OpenCode Repurpose rows come to 2.01 and 2.12 dollars per run. These are the authors’ own estimates, not an independent measurement.

What to take from it

The headline gain comes from the authors’ own runs on AgenticVBench, and the same-model design and ablations make it more informative than a cross-model comparison. Teams weighing the harness should read the cost table alongside the score, since the cheapest configuration and the most expensive one differ by a wide margin per run.

How we built this

Published results

Table 8 of the paper: Token usage and estimated cost by harness and task group

Rows of Table 8 of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column"; an empty name cell repeats the name above it. These are the authors' own figures, not an independent test.

arXiv

Sources

The sources this article was written from. Links open the original page.

  1. arXivVIVID: A Visual Harness for Agentic Video Editingarxiv.orgPrimary source · accessed November 17, 2026
  2. arXivVIVID: A Visual Harness for Agentic Video Editingarxiv.orgPrimary source · accessed November 17, 2026
  3. GitHubhttps://github.com/yuwenhan07/VIVID: page linked by the paper as its resourceraw.githubusercontent.comPrimary source · accessed November 17, 2026
  4. GitHubLicence of the repository yuwenhan07/VIVID on GitHubapi.github.comPrimary source · accessed November 17, 2026
  • video editing
  • agents
  • benchmark
  • open source
  • api cost