HeyGen just shipped a video model of its own, and the price is what got everyone talking. HeyGen Video sits on top of MiniMax H3, an open-weights model anyone can download, which HeyGen then retrained for its own needs. Rent it through the API and you'll pay as little as $0.01 a second until the end of October.
What Happened: HeyGen Built Its Own Model on Someone Else's Foundation
For the last couple of years, most AI video platforms worked the same way. They plugged in Kling, Veo, Seedance, MiniMax and half a dozen other models, wrapped a slick interface around them, and called it a platform. You were really just renting access to other people's engines.
Then MiniMax H3 arrived with open weights, meaning the model files are public and anyone can build on them, and the playbook changed. The new move is: take someone else's open foundation model, retrain it on your own data, tune it for your niche, and ship it as your own model.
HeyGen isn't pretending otherwise. The company says straight out that the model is built on MiniMax H3 and separately post-trained in-house. In plain English, H3 is the engine and HeyGen bolted on its own parts. The tuning targets consistency (does the same character or product still look right in shot five as in shot one), prompt following, product and reference handling, synced audio, and, in their words, more predictable commercial video production.
What HeyGen Video Can Do
• Text-to-video: type a description, get a clip.
• Image-to-video: turn a still photo into motion.
• Reference-to-video: feed it examples and tell it what to keep.
• Audio generated with the video, not stitched on afterward.
• Up to 9 images, 3 video clips and 3 audio files as references.
• You can call those references inside the prompt by name.
• Aspect ratios from 21:9 (cinematic) all the way to 9:16 (vertical phone video).
• 480p and 768p output.
• Clips run 5 to 15 seconds.
That last point matters. The pitch isn't about making viral slop fast. It's about control and stability, getting the same product, the same face and the same voice to hold up across multiple takes.
The Price Is the Real Headline
Until the end of October:
• 480p text-to-video and image-to-video: $0.01 per second
• 768p: $0.015 per second
• Reference-to-video at 768p: $0.03 per second
Do the math on a finished 10-second clip at 768p: 15 cents. That's less than a gumball.
For comparison, MiniMax's own H3 API runs roughly $0.08 per second at 768p, about 80 cents for the same ten seconds. So HeyGen fine-tuned the model and is now underselling the original. Third-party write-ups, including this launch overview and news coverage of the release, confirm the same specs and the same promo pricing.
The party doesn't last forever. After the promo, the cheapest 480p tier goes to $0.02 a second, and the premium reference-to-video tier comes in at $0.30 a second. Still cheap by 2024 standards, but several times the launch price.
HeyGen Says It Beat Seedance 2.0, H3 Max and Kling 3.0
On HeyGen's own blind evaluation, its version came out ahead of Seedance 2.0, H3 Max and Kling 3.0. Blind eval means the testers didn't know which model made which clip.
Take it with a grain of salt. It's the vendor grading its own homework, and nobody outside the company has reproduced it yet. The trend underneath is more interesting than the scoreboard, though: open weights are finally doing what they were always supposed to do.
What It Means for You
Think of it like a restaurant that used to resell another chef's dishes. Now the same restaurant takes a public recipe, tweaks the seasoning and sells the meal for less than the original chef charges. You're the diner, and dinner just got cheaper.
Here's what that looks like in practice.
At Home
You've got a folder of phone photos and one short story to tell. Image-to-video at 480p costs about a penny a second, so a 10-second clip with sound lands around 10 cents. Birthday montages stop being a weekend project.
At Work
A marketing team prepping a product launch. You upload up to 9 product shots, name them in the prompt, and generate variants that keep the label, the color and the logo consistent across every take. No reshooting, no studio rental.
For Business
An e-commerce seller needs one ad in three shapes: square for feeds, 9:16 for Reels and Shorts, 21:9 for a website banner. Same references, same look, three exports. At promo prices, the whole set costs less than a coffee.
For Students and Teachers
Reference-to-video is genuinely useful for explainers. Hand it a diagram, a photo and a voiceover file, then ask for a clip that keeps the diagram on screen while the narration plays. Five to fifteen seconds is exactly the right length for one idea.
For Creators and Freelancers
This is the income angle. Anyone selling short-form video to local businesses can now produce a draft in a couple of minutes for pennies and charge for the parts a model still can't do: taste, scripting and revision rounds.
If you'd rather test concepts before you spend anything, MyKreaTool gathers free AI tools you can use for scripts, thumbnails and reference images first.
How to Try It Right Now
Step 1: Check what's free. Log into HeyGen and read the pricing page before you load a card. Free option first, always. See what trial credits your account gets.
Step 2: Open the model page. Head to HeyGen Video and skim the input options. The model catalog page lists what the model supports.
Step 3: Pick your mode. Text-to-video if you're brainstorming, image-to-video if you have a product photo, reference-to-video if you need consistency across takes.
Step 4: Upload references. Up to 9 images, 3 videos and 3 audio files. Give each one a name you'll recognize, because you'll need it in the next step.
Step 5: Write the prompt and call your references by name. Something like, use hero_shot as the opening frame and keep the label readable. Plain language beats keyword soup.
Step 6: Set format and resolution. Choose your aspect ratio, then start at 480p for drafts, the $0.01 a second tier. Ten test clips cost you a dime.
Step 7: Keep clips between 5 and 15 seconds, then re-render the winners at 768p. That's $0.015 a second, so a finished 10-second spot runs about 15 cents.
Upsides and What Changes
The price floor just dropped. A second of decent AI video with synced audio is in penny territory now, and that changes what's worth making. You can afford to throw away 19 bad takes.
Audio is baked in. No separate voiceover pass, no lip-sync tool bolted on for every clip.
Consistency is the selling point. For product and brand work, that's worth more than raw photorealism.
Open weights are finally paying off for users. A foundation model went out the door, someone improved it, and the improved version is cheaper than the original API. That's the whole promise of open source, and it's happening in video right now.
The we-have-our-own-model era has started. Expect every AI video platform to announce a proprietary model in the next few months. Some will be real fine-tunes. Some will be a rebrand and a press release.
Limitations
Be clear-eyed about what you're actually buying. This is a closed product built on an open model, so you rent it rather than own it, and you can't run it on your own hardware. The only evidence that it beats Seedance 2.0, H3 Max and Kling 3.0 is HeyGen's own blind evaluation, so treat that ranking as a marketing claim until an outside party reproduces it. Output tops out at 768p, which is fine for social but not for broadcast or film, and clips cap at 15 seconds, so anything longer means stitching. The promo pricing expires at the end of October, and the fine print moves the cheapest tier to $0.02 a second. Also worth noting: reference-to-video at 768p costs twice as much as plain 768p, which tells you exactly where the real compute cost lives. Consistency isn't free.
Conclusion: One Action for Today
The interesting part of this launch isn't the leaderboard. It's the structure: an open model, retrained, sold cheaper than the original API, at a penny a second.
Your one action today: open HeyGen Video, check your free credits, and generate one 5-second clip at 480p from a single product photo or a photo of someone you know. At promo pricing, that costs about a nickel. You'll learn more in five minutes than from any benchmark chart.
Comments 0