Code2Video-AA ranks 27 models on writing videos in code: Claude Opus 5.5 leads, GPT-6.1 Sol is the value pick

On Oct 26 at 4:22 PM MDT, Artificial Analysis launched Code2Video-AA, built with HeyGen. The setup, per its methodology page: each model gets one design brief plus HeyGen's system prompt in a single call with no tools, and returns one HyperFrames HTML file (HTML for layout, GSAP for motion). Artificial Analysis renders it offline, and a Gemini 4 Argon judge compares two videos for the same brief without seeing model names. It scores five axes separately: engagement, prompt intent, composition, temporal consistency and craft. More than 35,000 verdicts become Elo ratings anchored to GPT-5.5 at 1000. The post passed 350 likes and 37K views overnight, and HyperFrames is on GitHub's TypeScript trending list today.

The results. Claude Opus 5.5 leads at 1536 and is first on all five axes. Its biggest lead is on craft and its smallest on prompt intent. Claude Sonnet 5.5 (1443) and GPT-6 Astra (1427) are statistically tied for second, and GPT-6.1 Sol is fourth at 1369. Cost per task ranges from $0.01 to $3.89 by Artificial Analysis's count: Opus costs $3.08 per task and Sol $0.31, about a tenth of that. The best open-weights models (DeepSeek V4.1 Flash, MiMo-V2.6-Flash and Pro, all around 1052) trail the leader by at least 483 points but cost $0.01–0.08 per task. Nearly everything rendered (98.8% of 3,078 tasks), so the gaps come mostly from the judge's taste, not from broken files.

Read it with the caveats in mind. It's a single shot with no tools, so it measures a model's first draft, not an agent that renders, watches and fixes its own work. The judge is also a model, Gemini 4 Argon, which itself places sixth. And every brief is a marketing video in HyperFrames, so Remotion or three.js work could rank differently.

The open-model side is moving too. SenseNova released SenseNova 6.8 Code2Video on Oct 24, a 27B Apache-2.0 fine-tune of Qwen3.8-27B with 512K context. It ships with a code-video skill for Claude Code-style harnesses, prompt enhancement and training trajectories. Its README says plainly that it is 'clearly behind the strongest proprietary models' on polish, that a production run takes hours, and that it reports qualitative examples only. It isn't on the leaderboard yet. Separately, OpenFilm (MIT, about 3,100 stars) takes a frameworkless approach: each shot is a web page drawn by a frame(t) function, with a local Studio editor for hand fixes afterwards.

HyperFrames kept shipping during all this. v0.8.207 to v0.8.209 (Oct 26–27) added keyframe interval and peak-bitrate controls, resumed stalled captures, fixed video after a cut running about 70 ms late, matched HDR preview to export, and handled phone videos whose first audio track won't decode.

Our take: if you pay for one model to draft motion graphics, this is the most useful public evidence so far. Use Opus 5.5 for hero pieces and GPT-6.1 Sol for volume, then let your agent's render-and-review loop close the rest of the gap. Treat the open-weights scores as a starting point, not a ceiling, since nothing with tools has been measured yet.