Introduction

HeyGen just stepped out from behind its avatars. HeyGen Video 1 (heygen-video-1) is the company's first general-purpose video model. It's a closed, API-first model built on MiniMax H3 and post-trained by HeyGen, and it creates the subject, setting, light and soundtrack of a shot in a single call. HeyGen announced it on X at 20:20 UTC on October 21 (04:20 MYT Oct 22) as "live on HeyGen + API". fal put three day-0 endpoints live about 40 minutes later.

Then, at 17:45 UTC on October 22 (01:45 MYT Oct 23), HeyGen posted that Video 1 ranked #1 in text to video in Artificial Analysis' blind arena. That post passed 1.4 million views. The fine print matters, so here's exactly what shipped, where the #1 claim comes from, and how to try it today.

What Shipped

From HeyGen's model page and developer docs:

  • Three modes, one model: text_to_video, image_to_video (your image becomes the literal first frame) and reference_to_video (it keeps products, people or places you supply).
  • Sound in the same pass: every clip comes with its own dialogue, room tone and effects, so you don't need a separate voiceover or lip-sync step. Output is 24 fps H.264 MP4 with AAC 32 kHz stereo audio.
  • Resolutions: 480p, 768p, 1080p and 2K. On HeyGen's API, 1080p and 2K bill at 3× the 768p rate, and HeyGen's model page says they render in 16:9 or 9:16 only.
  • Length: a new October changelog entry raises text-to-video and image-to-video to 5–30 seconds with turbo or quality prompt enhancement. Reference-to-video, ultraquality, disabled enhancement and the HeyGen MCP stay at 5–15 seconds.
  • References: up to 9 images, 3 videos and 3 audio clips, addressed in the prompt as <Picture 1>, <Video 1>, <Audio 1>. Only the first 5 seconds of each reference video are read.
  • Prompt length: up to 32,000 characters. HeyGen says long, detailed prompts beat short ones, and a fixed seed returns the same clip, so you can change one clause at a time.

HeyGen says the model works best on short, contained shots with one subject, one place and one action. It's least reliable on long on-screen text, soft organic motion like petals, paper and hair, and close hand work.

The Arena Fine Print

HeyGen's main account says "#1 in text to video." Its developer account adds the detail. In Artificial Analysis' blind human votes, HeyGen Video 1 (benchmarked with ultraquality prompt enhancement) is #1 on Text to Video (silent) and #3 on Text to Video with audio.

The Artificial Analysis leaderboard backs that up. Here's the snapshot when we checked it around 23:00 MYT Oct 23:

  • Silent T2V: HeyGen Video 1 at Elo 1144, ahead of Wan 3.0 (1134) and MiniMax H3 768p (1120).
  • T2V with audio: Wan 3.0 (1175), Utopai X (1167), then HeyGen Video 1 (1160), just above MiniMax H3 768p and Dreamina Seedance 2.5 (both 1158).

So the post-trained H3 now edges out its own base model on both boards. Treat the gaps as narrow, though: Artificial Analysis lists HeyGen's confidence interval at about ±17 Elo on the silent board and ±15 with audio, so the top spots overlap. Also keep in mind that the ranked setting (ultraquality) is one of the settings capped at 15 seconds.

Where to Run It Today

  • HeyGen API: POST https://api.heygen.com/v3/models/videos with an x-api-key. HeyGen's model page lists $0.01 per second as a launch price through October.
  • HeyGen app: listed as free to start on HeyGen's model page, in the same workspace as your avatars and voices.
  • fal: heygen/heygen-video/text-to-video, /image-to-video and /reference-to-video at $0.02/s for 480p and 768p and $0.06/s for 1080p and 2K, 5–15 seconds per clip, under a commercial license.
  • ComfyUI: the HeyGen Video 1.0 Image to Video and Reference to Video partner nodes already in core (2K added in early October), capped at 5–15 seconds.

Here's a 20-second text-to-video call shaped the way HeyGen's prompting guide suggests (name the camera, describe the sound, and rule out what you don't want):

curl -X POST https://api.heygen.com/v3/models/videos \
  -H "x-api-key: $HEYGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "heygen-video-1",
    "mode": "text_to_video",
    "prompt": "A night-market noodle cook in a plain black apron tosses noodles in a flaming wok under string lights, then plates them and slides the bowl toward camera. Shot on an ARRI Alexa Mini LF, 35mm T2, locked off at counter height. Sound: wok sizzle close, crowd murmur and distant scooter horns, no music. No logos, brand names or printed words anywhere in frame.",
    "duration": 20,
    "resolution": "1080p",
    "aspect_ratio": "16:9",
    "prompt_enhancement": "quality",
    "seed": 42
  }'

Then poll GET https://api.heygen.com/v3/models/videos/{video_id} until the status reads completed, and grab the signed video_url.

Prefer a paste-ready prompt for the app or fal? Try this one:

A violinist in a plain grey wool coat plays a slow phrase on an empty rooftop at blue hour, city lights coming on behind her, breath visible in the cold air. Shot on an ARRI Alexa Mini LF, 50mm T2, locked off on a tripod at chest height, shallow depth of field. Sound: solo violin close and warm, faint wind, distant traffic far below, no music bed. No retouching, no beauty filter, no logos, brand names or printed words anywhere in frame.

Conclusion

HeyGen Video 1 is a big swing from an avatar company. It's a post-trained MiniMax H3 that tops Artificial Analysis' silent text-to-video board, sits #3 once sound is judged, and costs a launch-week $0.01 a second on HeyGen's own API. If you make B-roll, product scenes or social hooks, October is the cheap window to stress-test it. Write long prompts, lock a seed, and say out loud what the scene should sound like.

Original Source

—Aurelia ♡