Skip to content
Advertisement
OpenAI’s GPT-6.2 Astra takes on ByteDance’s Dreamina Seedance 2.5 in a matched video test

OpenAI’s GPT-6.2 Astra takes on ByteDance’s Dreamina Seedance 2.5 in a matched video test

FoxTPNL / Wikimedia Commons (CC BY 4.0)

An Artificial Analysis comparison pits code-built 3D scenes against native AI video using identical camera paths

Two very different philosophies of AI video just went head to head. On November 20, 2026, Artificial Analysis ran an independent comparison of OpenAI’s GPT-6.2 Astra against ByteDance’s Dreamina Seedance 2.5, with both models working from identical camera paths.

The test used three scenes, each chosen to stress a different skill. The lineup: the Sydney Opera House, an exploded view of a mechanical watch, and a dragon attacking a medieval town.

GPT-6.2 Astra did not output video in the usual sense. It wrote three.js code, a JavaScript framework for building 3D environments, and the resulting scenes could be viewed interactively in a browser at 1080p and 24 fps.

Seedance 2.5 took the traditional route. It generated native video from text prompts, guided by reference images showing the camera routes it needed to follow.

Both models used the same camera paths and settings.

Meet the contenders

GPT-6.2 Astra is fresh out of the box. OpenAI released it on November 17, 2026, just three days before the test, positioning it as a flagship model.

Advertisement

OpenAI’s pitch for Astra centers on complex workflows, coding, and constructing 3D scenes.

Seedance 2.5 is a specialist with a cinematic focus. It can generate videos of up to 30 seconds in a single pass, and it accepts up to 50 reference assets per request.

Through its API, Seedance 2.5 costs roughly $0.23 per video second.

As of mid-November 2026, Seedance 2.5 held the #3 spot on Artificial Analysis’s Text-to-Video Elo leaderboard.

Elo, borrowed from chess rankings, scores models by how often they win head-to-head matchups.

The architect versus the cinematographer

The test highlights a split emerging in AI video. According to the analysis, language models are increasingly handling the geometric side of video, while dedicated video models focus on rendering and visual coherence.

Astra plays architect. Because its output is code, the geometry is explicit: objects have defined positions, and the camera follows a path through actual 3D space.

Seedance plays cinematographer. Native video models generate pixels directly, which tends to favor texture, lighting, and atmosphere.

What this means for creators and the AI race

For creative teams, the two approaches solve different problems. A browser-based 3D scene from Astra is editable and explorable after the fact, while a Seedance clip is a finished piece of footage.

The pricing models diverge too. Seedance charges by the video second, so costs scale with runtime. Code-generated scenes, by contrast, are produced once and then rendered on the viewer’s own machine.

For ByteDance, Seedance 2.5’s long single-pass clips and heavy reference support reflect a bet on professional production.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.