Code2Video-AA: Claude Opus 5.5 Leads New Video-Design Benchmark at 1536 Elo, 93 Points Clear
Artificial Analysis published Code2Video-AA on October 26, 2026: an independent implementation of HeyGen’s Code2Video Bench, measuring how well AI models design short product and brand videos by writing code rather than prompting a video generator.
How the benchmark works
Each model receives one design brief and HeyGen’s system prompt in a single call, with no tools. It returns one HyperFrames HTML composition — HTML for layout, GSAP for animation — which is then rendered to video. Five judge axes score each video: craft, prompt intent, temporal consistency, and two additional dimensions. The overall Elo is the mean across all five. Comparisons are sampled pairwise and fit by standard Elo methods. The benchmark runs 114 briefs from the Code2Video Bench public split, licensed CC BY 4.0.
Results
Claude Opus 5.5 scores 1536 overall Elo, anchored against GPT-5.5 (xhigh) at 1000. Second place is a near-tie: Claude Sonnet 5.5 at 1443 and GPT-6 Astra at 1427 are within noise. GPT-6.1 Sol places fourth at 1369.
| Model | Elo |
|---|---|
| Claude Opus 5.5 (max) | 1536 |
| Claude Sonnet 5.5 (max) | 1443 |
| GPT-6 Astra (max) | 1427 |
| GPT-6.1 Sol (max) | 1369 |
| Claude Fable 5.1 (max) | 1325 |
| Gemini 4 Argon (high) | 1259 |
| Claude Haiku 5.5 (max) | 1192 |
| Grok 4.7 (xhigh) | 1127 |
| DeepSeek V4.1 Flash (max) | 1053 |
Claude Opus 5.5 leads on all five judge axes. The margin is largest on craft (128 points over Claude Sonnet 5.5) and smallest on prompt intent (48 points). Temporal consistency shows the least separation across models: its Elo range is 1057 points, against 1222–1349 points on the other four axes.
Cost gap
Cost per task ranges from $0.01 to $3.89 — more than 300x across the tested models. Claude Opus 5.5 costs $3.08 per task. GPT-6.1 Sol, at $0.31, ranks fourth while costing roughly one-tenth as much. That makes Sol the clearest value play: fourth-best output at sub-cent-per-task territory.
Open weights
The top open-weight models — DeepSeek V4.1 Flash, MiMo-V2.6-Flash, and MiMo-V2.6-Pro — cluster at 1052–1053 Elo, 483 points below Claude Opus 5.5. Every open-weight model sits below 1060. The gap holds across all five judge axes: open models can render the videos (render rates range from 96% to 100%), but the quality ceiling is materially lower.
Render rates
3042 of 3078 tasks (98.8%) rendered successfully. Eight models rendered all 114. GLM-5.3 has the lowest render rate at 110 of 114. The three lowest-ranked models rendered 111–113 briefs each, so their low Elo reflects judge verdicts on the videos that did render, not failure to produce output.
What it measures
Code2Video-AA is a narrow task: one brief, one call, one rendered video. It does not test multi-turn refinement, iteration, or tool use. The 300x cost spread and the open-weights gap suggest that video layout composition in code is a domain where frontier model capacity — not scaffolding or prompting strategy — is still the primary variable.