Megaton H2H / Creative prompting
Four prompts. One Odyssey.
Frontier language models wrote competing production prompts for the same brief. Seedance 2 rendered each prompt unchanged, and Megaton scored the results.
OpenAI
GPT 5.6 Sol
65%
Anthropic
Claude Fable 5
75%
Moonshot AI
Kimi K3
72%
Anthropic
Claude Opus 5
73%
The outputs
Watch all four generations
All clips are eight seconds, 720p, 16:9, with native audio.
OpenAI
GPT 5.6 Sol
65%
Anthropic
Claude Fable 5
75%
Moonshot AI
Kimi K3
72%
Anthropic
Claude Opus 5
73%
All pairings
Open a comparison
Scoring
A simple rubric
One overall human score, guided by four equally weighted questions. The percentages are not model-generated.
- 01
Prompt adherence
Does the video deliver the requested actions, sequence, and historical details?
- 02
Visual craft
Are the composition, cel-anime aesthetic, lighting, and material details convincing?
- 03
Motion and continuity
Do characters, props, geography, and transitions remain readable and stable?
- 04
Story and sound
Does the eight-second arc communicate clearly, with audio that supports the scene?
Shared brief
The Odyssey: Sirens
Create an eight-second, 16:9, hand-drawn 2D cel-anime retelling of Odysseus and the Sirens in exactly four connected shots. Use historically grounded Late Bronze Age Aegean materials, precise character action, cinematic camera direction, and native sound.
Each frontier model received the same creative brief through Vercel AI Gateway when available, with a direct provider fallback.
Each response was passed to Seedance 2.0 without prompt editing.
Every output used the same 720p, 16:9, eight-second settings with native audio.
The percentages are human editorial scores guided by the rubric below.