How Code2Video-AA Works: Single-Shot Generation, Blind Pairwise Elo Judging
ArtificialAnlys · x · 2026-10-27
Code2Video-AA methodology: each model reads one design brief and writes a single HyperFrames HTML composition with no tools; compositions are rendered offline, and a failed render loses all comparisons. A Gemini 4 Argon judge watches pairs of videos at 8 fps and picks the better one per axis across five blind comparisons. Bradley-Terry fits per axis yield Elo ratings, averaged into an overall score anchored at GPT-5.5 (xhigh)=1000. The judge itself ranks 6th on the leaderboard at 1259.
More from Models
- Apple open-sources Private Cloud Compute components, giving researchers a real attack surface — jedisct1 · 2026-10-27
- DeepSeek-V4.1-Flash-1017 hits original's max-effort score at default effort: $0.15 per correct task — ryanshrout · 2026-10-27
- Artificial Analysis teases Intelligence Index v5 launch on October 27 — ArtificialAnlys · 2026-10-27
- Anthropic's Blended Token Price Nearly Halved to $0.80/M in 10 Months, Ramp Data Shows — FinanceYF5 · 2026-10-27
- GPT-6.1 Terra Is the Cheapest Frontier Model at $0.67 per Successful Task — bindureddy · 2026-10-27
- Nemotron-3.5 Bolt hits 9B active params, within 1% of Super on agentic benchmark — ryanshrout · 2026-10-27