Claude Opus 5.5 wins Code2Video-AA with 1536 Elo, beating GPT-6 Astra by 109

Artificial Analysis has released Code2Video-AA, an independent implementation of HeyGen's Code2Video Bench. Models receive 114 design briefs for short product and brand videos and must build each video in code. In the first ranking, dated October 26, 2026, Claude Opus 5.5 (max with fallback) leads with 1536 Elo. It is 93 points ahead of second place and ranks first on all five evaluation axes. Claude Sonnet 5.5 (1443) and GPT-6 Astra (1427) follow, and the gap between them is within the margin of noise.

How a model makes video without a single tool

The conditions are strict. The model gets the brief and HeyGen's system prompt in a single call, with no tools and no chance to fix the result. The output must be a single index.html file containing a HyperFrames composition: HTML handles layout and GSAP handles animation. The composition is then rendered to video, and LLM judges compare the clips pairwise on five axes. The final Elo is the average across the axes, anchored so that GPT-5.5 (xhigh) sits at 1000.

The briefs are highly detailed. One example from the public sample asks for a wireframe-builder demo of about 7.3 seconds. It specifies a black title card, a hard cut, a static camera with no zoom, a strict palette with a single lime-yellow accent, #E4E41B, and Spacer, Media and Text blocks that appear on the canvas one after another under a cursor. The model has to hold the timing and the palette in code.

Table: 27 models by Elo

• Claude Opus 5.5 (max with fallback) — 1536
• Claude Sonnet 5.5 (max with fallback) — 1443
• GPT-6 Astra (max) — 1427
• GPT-6.1 Sol (max) — 1369
• Claude Fable 5.1 (max with fallback) — 1325
• Gemini 4 Argon (high) — 1259
• Claude Haiku 5.5 (max) — 1192
• Grok 4.7 (xhigh) — 1127
• DeepSeek V4.1 Flash (max) — 1053
• MiMo-V2.6-Flash — 1052
• MiMo-V2.6-Pro — 1052
• Kimi K3 (max) — 1041
• GLM-5.3 (max) — 1034
• GPT-6 Luna (max) — 1030
• Qwen3.8 Max (0902) — 1021
• Muse Spark 1.3 (max) — 1019
• GPT-5.5 (xhigh) — 1000
• Muse Wonder (max) — 997
• GLM-5.3-Flash — 975
• Qwen3.8 27B (xhigh) — 949
• Mistral Large 4 Preview — 931
• Gemini 3.8 Flash (high) — 899
• Step 5 Preview — 821
• MiniMax-M3 — 724
• Beam 1.0 — 675
• Inkling (xhigh) — 467
• Nemotron 3 Ultra — 304

The gap between the leader and the open models is at least 483 points. The best open models are DeepSeek V4.1 Flash (1053) and MiMo-V2.6-Flash and MiMo-V2.6-Pro (1052 each).

Opus 5.5's advantage is most visible in craft, meaning cleanness of execution, where it beats Sonnet 5.5 by 128 points. The smallest lead is on prompt intent, at 48 points. The model order barely changes from axis to axis: no model moves more than six places. Temporal consistency separates the models least: the Elo range there is 1057 points, against 1222–1349 on the other four axes.

The price of winning and token consumption

The cost of a single task ranges from $0.01 for MiMo-V2.6-Flash to $3.89 for Claude Fable 5.1, a difference of more than 300x. The leader costs $3.08. GPT-6.1 Sol, in fourth place, costs $0.31, roughly ten times less. GPT-6 Astra costs $1.62, 5 times more than Sol for a gain of 58 Elo points. Eight models sit on the Pareto front of Elo versus price. They include MiMo-V2.6-Flash (1052 Elo, $0.01), Claude Haiku 5.5 (1192 Elo, $0.08), GPT-6.1 Sol (1369 Elo, $0.31) and Claude Opus 5.5 (1536 Elo, $3.08).

Token consumption differs 16-fold: from about 11 thousand for Nemotron 3 Ultra to about 176 thousand for Claude Sonnet 5.5. Most of the difference comes from reasoning tokens. The responses containing the compositions themselves take 4 to 12 thousand tokens for every model. The three Claude 5.5 models spend 152–176 thousand tokens per task. The other 24 models average about 42 thousand, with a median of about 34 thousand.

A separate check covers rendering: a composition that fails to turn into a video loses every comparison. 3042 of 3078 tasks rendered (98.8%), and 8 of the 27 models completed all 114. The lowest rate belongs to GLM-5.3 (max): 110 of 114. The bottom three models rendered 111–113 tasks, so their low Elo comes mostly from the judges' verdicts on finished videos.

The practical takeaway is this. If you need the best code-driven animation quality and the budget allows it, Opus 5.5 has no competitors for now. If cost matters across a high volume of videos, GPT-6.1 Sol or Claude Haiku 5.5 are the more sensible options. Open models are so far good only for cheap drafts: MiMo-V2.6-Flash costs pennies but trails by 484 points. The ranking will be updated on the Code2Video-AA page, and HeyGen has published the brief sample under a CC BY 4.0 license.