What shipped
HeyGen released its first general-purpose video model. It is built on MiniMax H3 and post-trained in-house, and it handles text-to-video, image-to-video and reference-to-video in a single model with synchronised sound.
Coming from HeyGen, the obvious read is “avatar company expands into general video”. That is true but less interesting than the second-order point: the company that has spent years on lipsync and presenter video is now the one shipping synchronised audio as a first-class output rather than an afterthought.
Three input modes in one model
Most tools specialise. You pick a text-to-video tool for concepting and a different image-to-video tool for animating existing assets. Collapsing that into one model has real workflow value.
| Mode | Input | Best for |
|---|---|---|
| Text to video | A written prompt | Concepting, B-roll, shots that do not exist yet |
| Image to video | A still image | Animating product photography, artwork, key frames |
| Reference to video | Reference images of a subject | Keeping a person or product recognisable across shots |
Reference-to-video is the one to watch: it is the mode that addresses character and product consistency, which remains the single biggest production complaint about AI video. Whether it delivers at scale is the open question.
Why synchronised sound matters
Almost every video model outputs silent footage. Sound is then added in a separate pass — voiceover recorded or generated, music licensed, effects placed, everything aligned by hand. That pipeline is where a lot of the real production time goes.
- Synchronised output removes an entire alignment step for dialogue and presenter content.
- It matters most for the use case HeyGen already dominates: talking-presenter and avatar video.
- For cinematic or music-led work, silent generation plus a proper audio pass is still the better route.
What one cent per second actually means
The launch rate is roughly one cent per second, about half the standard price. Worked through, that is around sixty cents for a minute of raw output — but raw output is not finished output.
| Measure | At promotional rate | Reality check |
|---|---|---|
| One minute generated | About $0.60 | Assumes every second is usable |
| One minute with 3x regeneration | About $1.80 | Closer to typical for demanding shots |
| Finished minute after editing | Higher | Add your own edit and audio time |
Promotional pricing is a subsidy: it exists to win evaluation, not to describe your long-run cost. Model on the standard rate and treat the discount as temporary unless contractually locked.
Why post-training on H3 is the real story
Building on an existing strong base model and post-training in-house is now the dominant pattern, and it tells you where the market is.
- Base models are commoditising. Several strong open and closed bases exist, and training one from scratch is increasingly hard to justify.
- Differentiation is moving to post-training on proprietary data. What you fine-tune on becomes the moat.
- Specialisation beats generality at the margin. A model post-trained on ad work outperforms a general model on ad work, even from the same base.
This is why you see the same base model appearing under several product names with genuinely different behaviour. The base matters less than the post-training data.
Who this suits
- Presenter-led content: training, internal comms, product explainers where a person talks to camera.
- Localisation at scale: the same script delivered in multiple languages with matched delivery.
- Budget-constrained enterprise teams that need volume more than cinematic range.
- Teams already inside HeyGen, where keeping the workflow in one place is worth more than marginal quality gains elsewhere.
It is a weaker fit for cinematic work, music-led edits, or anything where you need precise control over camera and motion rather than a presenter delivering a script.
What to verify before committing
-
Test reference-to-video properly
Generate the same subject across five shots and check whether it stays recognisable. This is the claim most likely to matter to you and the one most likely to disappoint.
-
Check audio quality on your real scripts
Synchronised sound is only useful if the voice is usable. Test with your actual scripts, including numbers and proper nouns.
-
Confirm the rate end date
Ask directly what happens at the end of the promotional period and whether existing customers are grandfathered.
-
Verify commercial terms
Promotional tiers sometimes carry usage restrictions. Confirm commercial rights in writing before shipping client work.