HeyGen Charges a Penny a Second, and the Video Model Price Floor Just Moved

HeyGen released HeyGen Video this month, its first general-purpose video model. It is built on MiniMax H3 and post-trained in-house, covering text-to-video, image-to-video and reference-to-video in one model with synchronized sound. The promotional rate at launch is $0.01 per second, half the standard price, aimed squarely at budget-constrained enterprise use.

That number is worth sitting with. A ten-second clip costs a dime. Last year the same category of output was priced in dollars per generation, and teams rationed it.

The promotional framing is itself informative. HeyGen set the standard rate at two cents a second and launched at half that, which is a price cut aimed at pulling teams off incumbent vendors rather than at maximizing revenue per clip. A promotional rate is a bet that once a workflow is built around a model, the switching cost keeps the customer after the discount ends.

The per-second rate is the new benchmark

Video model pricing has moved from per-generation to per-second of output, and the published rates now span an order of magnitude. HeyGen's promotional penny per second converts to about $0.60 per minute. MiniMax H3, the open-weight base, runs at $4.80 per minute. Wan 3.0 is listed at $12 per minute. Seedance 2.5 charges $34.12 per minute. Grok Imagine Video 1.5 sits at $0.08 per second, or $4.80 per minute, with native audio and a single published rate with no resolution tier.

Grok's single-rate structure is worth noting on its own. Most vendors charge by resolution tier, so a 1080p render costs more than a 480p one. One published rate with no tier is easier to budget against and harder to game. It also signals confidence that the model's quality does not depend heavily on how many pixels the buyer requests.

The spread tells you the market has segmented. At the top is quality for campaigns where a single clip matters. At the bottom is volume for workflows where a hundred clips get generated and most get discarded.

HeyGen's pricing puts it deliberately at the low end, and the framing is enterprise use cases where the budget is the constraint. That is a different customer than an agency producing a hero spot. It is a company that wants video in every support article, every product page, every onboarding flow, and cannot justify a designer for each one.

Post-training on someone else's base

The architecture choice is as notable as the price. HeyGen did not train a foundation model from scratch. It took MiniMax H3, an open-weight model, and post-trained it in-house for its use case.

This is becoming the standard playbook. Creatify did the same thing with Boreal-H3, post-training MiniMax H3 on real ad-project data and reporting brief completion rising from 27.8% to 50% and subject consistency from 83.3% to 94.4% against the base. The pattern repeats because the base is open and the differentiation lives in the training data and the evaluation, not in the pretraining run.

For teams building video products, the calculation is straightforward. Training a foundation model costs tens of millions and takes months. Post-training an open base costs a fraction of that and produces a model tuned to your specific quality bar. The risk is that your product's core capability sits on someone else's license terms, which is a live issue for H3, since MiniMax has not published which territories its license covers. Three products built on H3 launched this week in territories the license may not reach, and the lab welcomed them, which is a long way from granting permission.

The license question is the part of the post-training playbook that gets the least attention. A model can be downloaded today and unavailable tomorrow, or legal in one market and not another, and the terms are often written for the model's original purpose rather than for the products built on top of it. Anyone whose product depends on an open base should read the license as carefully as they read the technical report. The cost advantage of post-training is real, and it comes with an operational risk that does not show up in a pricing comparison.

What the price means for creative work

When a ten-second clip costs a dime, the economics of iteration change. A team can generate fifty variants of a product demo, watch them, keep the two that work, and spend five dollars doing it. The bottleneck moves from generation cost to human review time, which is the opposite of where it sat two years ago.

That shift has consequences for how video teams are staffed and how workflows are designed. Cheap generation means the expensive part is deciding what is good, and the people who can make that judgment quickly become the constraint. It also means the flood of mediocre output gets larger, because the cost of producing something no one needed just fell to near zero.

There is a less obvious effect on how teams plan. When a clip costs a dime, the rational strategy shifts from making one good version to making many and selecting. That favors teams who are good at evaluation and disfavors teams who are good at production craft, because the craft that used to be the value is now a commodity input.

It also changes what a client expects. A brief that used to specify one video now specifies five variants for A/B testing, because the marginal cost of the extra four is negligible. The work moves upstream to strategy and downstream to analytics, and the middle, where most creative production labor sits, thins out.

The question the price does not answer

Cheap generation is not the same as useful generation. A penny a second buys a clip, not a clip that fits the brief. The differences that matter in production, keeping a product label readable, keeping a creator's face consistent from first frame to last, following a script rather than a vibe, are the things the higher-priced tiers charge for.

That is the real segmentation the price spread reveals. The market is separating into models for volume and models for precision, and the price difference between them is roughly a factor of fifty. A team that knows which category its work falls into can make a clean decision. A team that assumes the cheap model will do can discover the gap on a client deliverable.

HeyGen's promotional rate will not last at a penny. What it does is set a floor the rest of the market has to answer, and force everyone building video features to model their unit economics at a price point that would have been implausible last year. The interesting question is which capabilities turn out to be worth paying more than a penny a second for.

The answer is likely to be narrow and specific. Label legibility, identity consistency and brief adherence look unglamorous, but they are the ones that decide whether a clip can be shipped. Everything else is a matter of taste, and taste is cheap to iterate on when a take costs a dime.