Three models made the same 8 second Doe hook from one bare prompt, 2026-11-08

Each model got the same one-sentence prompt and no voice sample, so each one made its own voice and its own woman.

One transcript heard all 24 words in all three takes, while the stricter QA read heard Seedance say "dermologists" for "dermatologists".

The players start muted, so turn the sound on to judge the voices.

Seedance 2.0 mini cost $0.28 and her head moved 6.2% of her face width

She holds a second phone in frame at the start, and QA marks the slurred "dermologists" as a major defect.

MiniMax H3 Max on fal cost about $0.40 and her head moved 3.8% of her face width

She gestures with her hands while she talks, and QA marks her lips as running at least 333 ms off the sound, which is a major defect.

Base H3 on our rented GPU cost $0.43 with its setup and her head moved 5.2% of her face width

She gestures with her hands, and QA marks a 0.6 second pause after "sentence" as a major defect.