AI-Drawn Video Model Leaderboard
Compare 14 author models drawing the same six briefs with Ruvil 2.1. Start with the technical all-pass percentage, then compare recorded spend, billed tokens and elapsed time.
Model leaderboard
Sorted by best %
| Rank | Model | Score % | Spend | Cost/film | Tokens/film | Wall time |
|---|---|---|---|---|---|---|
| 1 | Muse Spark 1.3 (spark-1-3); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $3.31 | $0.5513 ± 0.0138 | 275,657 | 5.53 ± 0.16 min |
| 1 | MiMo V2.6 Pro (mimo-v2.6-pro); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $3.63 | $0.6053 ± 0.0253 | 302,640 | 46.38 ± 5.80 min |
| 1 | Gemini 3.8 Flash (gemini-3-8-flash); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $4.77 | $0.7957 ± 0.0842 | 397,841 | 10.83 ± 1.32 min |
| 1 | Qwen3.8 Max (qwen-max); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $5.73 | $0.9548 ± 0.0420 | 477,381 | 18.44 ± 1.38 min |
| 1 | GPT-6.1 Sol (sol-6-1); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $8.18 | $1.3629 ± 0.1301 | 681,461 | 20.31 ± 1.70 min |
| 1 | Gemini 4 Argon (gemini-4-argon); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $11.85 | $1.9751 ± 0.1776 | 987,563 | 18.96 ± 1.29 min |
| 1 | Claude Opus 5.5 (opus-5-5); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $12.63 | $2.1051 ± 0.1750 | 1,052,542 | 22.21 ± 10.96 min |
| 1 | Muse Wonder (muse-wonder); 6 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $13.46 | $2.2433 ± 0.1087 | 1,121,645 | 11.50 ± 0.53 min |
| 1 | GLM 5.3 (glm-5.3); 5 eligible, 6/6 delivered | 100.00%; SEM 0.00 pp | $6.12 | $1.0192 ± 0.2071 | 509,585 | 21.79 ± 6.89 min |
| 1 | Claude Fable 5.1 (fable-5-1); 1 eligible, 2/6 delivered | 100.00%; SEM N/A | $10.04 (provisional) | $1.6727 ± 0.6958 | 836,355 | 22.08 ± 3.90 min |
| 11 | DeepSeek V4.1 Flash (deepseek); 6 eligible, 6/6 delivered | 83.33%; SEM 16.67 pp | $3.17 | $0.5277 ± 0.0576 | 263,836 | 17.15 ± 2.11 min |
| 11 | Kimi K3 (kimi-k3); 6 eligible, 6/6 delivered | 83.33%; SEM 16.67 pp | $12.23 | $2.0375 ± 0.4056 | 1,018,757 | 34.74 ± 7.29 min |
| — | Claude Sonnet 5.5 (sonnet-5-5-openrouter); 0 eligible, 2/6 delivered | N/A; SEM N/A | $23.73 | $3.9546 ± 1.1006 | 1,977,278 | 61.40 ± 0.02 min |
| — | GPT-6 Astra (astra-6); 0 eligible, 6/6 delivered | N/A; SEM N/A | $5.95 | $0.9918 ± 0.0504 | 495,888 | 38.15 ± 0.77 min |
Score is technical observable all-pass on eligible delivered films, not request success or artistic quality. Equal scores share a rank; ties are displayed by eligible sample size, then lower spend. N/A is not zero.
Spend is recorded billed North Noir tokens in USD ($2 per million), not provider cost. Cost/film and Tokens/film average all six submitted film requests, including recorded debits on non-deliveries; tokens are customer billing tokens, not LLM usage. Wall time averages deliveries only. Fable billing is provisional.
By Syver-August, Developer, North Noir · · 5 min read
What this measures
This is a technical-observable comparison, not an artistic-quality grade. Ruvil 2.1 runs a JavaScript drawing program for each silent 10-second film at 1920×1080 and 30 fps, with up to three product review rounds. The score is the share of eligible delivered films passing all three checks below; identity, anatomy, construction order, fill ownership, contact and editability are not graded. The independent benchmark jury made zero calls.
- Saved-frame contract: indices, hashes, dimensions and timing, plus the retained probe contract. This is not an every-frame PNG-to-MP4 linkage proof.
- Specific edge pattern: no detected instance of the published dark-on-light edge pattern. It is not a universal no-border guarantee.
- Exact-repeat run: no run of three or more decoded-RGB-identical frames outside declared holds. Different pixels alone do not establish adequate motion.
How it was run
grok-4.7, Claude Haiku 5.5 and MiniMax M3.1 Flash are not ranked because they completed too few runs.
Reading percentage and spend together
All-pass is conditional on eligible delivery, not success across all requests. Sample sizes sit next to the score: Fable has one eligible film, GLM has five, and Astra and Sonnet have no eligible author/source-verified film, so their percentages remain N/A. A tie does not imply equal artistic ability. SEM describes this small public-brief cohort, not a general confidence guarantee. Cost/film averages recorded debits over six submitted requests, not just deliveries; Fable’s recorded bill is provisional and lower-observed. The spend column is the sum of those debits, and Tokens/film converts the same North Noir billing at $2 per million customer tokens. Wall time includes authoring, product review, polling, downloading and analysis, not just CPU runtime.
Use the drawing workflow
The model writes the drawing program; you direct the result. Learn how Ruvil 2.1 draws and reviews each section, explore it in the Studio, or use the CLI with the AI coding harness you already have. Choose a model, inspect the frames and use Adjust to refine the details that matter to your own film.
Key takeaways
- Compare 14 models by technical all-pass percentage, with eligible sample sizes alongside each score; this is not an artistic-quality ranking.
- Spend, Cost/film and Tokens/film use recorded North Noir customer billing. They are not model-provider costs or LLM input/output usage.
- Equal scores share a rank. N/A remains unavailable rather than becoming zero, and the wall-time average covers delivered films only.
Frequently asked questions
What does Score % mean?
It is technical observable all-pass: the percentage of eligible delivered films passing the saved-frame contract, specific edge-pattern check and exact-repeat-run check together. It is not an artistic-quality score or the success rate across all requested films.
How is the leaderboard sorted?
By best technical percentage first, with unavailable percentages last. Equal scores share a rank; tied rows are displayed by larger eligible sample, then lower recorded spend. Click a column header, or use the mobile sort control, to compare another metric.
Why are Astra and Sonnet marked N/A?
Their delivered artifacts do not provide an eligible selected-author/source-verified film for the technical summary. Their recorded spend remains visible, but no percentage or numerical rank is assigned.
What is the difference between Spend and Cost/film?
Spend is the model’s total recorded customer-token bill across six requests. Cost/film divides that total by six, including recorded debits on requests without a delivery; Tokens/film uses the same denominator and is North Noir billing, not model token usage. Fable billing remains provisional.
Does this measure which model makes the best animation?
No. Semantic identity, anatomy, construction order, contact and editability are outside these technical checks. Use the numbers to compare the observed technical behaviour and billing, then inspect the drawings against your own creative brief.