On August 3, MiniMax surprise-launched H3, its next-generation multimodal generation model. What detonated the industry’s reaction wasn’t the model’s capability — it was the price sheet that collapsed over the following days to levels nobody had priced for.
The official rate is 0.8 RMB per second at 2K resolution — already two-thirds cheaper than mainstream flagships. Then third-party platform Metaso AI undercut that brutally: 0.09 RMB per second at 768P, and 0.15 RMB per second at 2K. Run the numbers: a 5-second 2K clip costs 0.75 RMB (about ten cents); 10 seconds costs 1.5 RMB — the price of a bottle of water. At 768P, five seconds comes in under 0.5 RMB.
This is not a price cut. It is a structural migration of the entire price band: AI video generation has formally moved from “pay-per-second heartache” into a pennies-per-clip era where experimentation doesn’t hurt.
How H3 Actually Cut the Price
The low price is not subsidized burn — it is architecture. During training, every frame of video gets chopped into small patches fed to the model; more patches mean costlier training and inference. H3’s core breakthrough is a proprietary VAE (H3-VAE) that pushes compression to an industry-leading ratio — the same footage expressed in fewer patches, cutting sequence length by a factor of four.
Resolution strategy also diverges from the mainstream. The prevailing approach generates a low-resolution version first, then bolts on a super-resolution module to upscale to 2K/4K. That module is essentially guessing the missing detail, which is why small text, textures, and lighting often blur. H3 refuses to treat the low-res output as the sole ground truth: it uses it as a reference alongside the original multimodal context, reusing H3’s existing generative capability to recover information invisible in the low-res video but still present in the context.
That is also the underlying reason H3 doesn’t fall apart on complex text scenes — titles, logos, subtitles. In one sentence: H3 competes on architectural efficiency, not on burning money for market share.
The Multimodal DeepSeek Moment
The script looks familiar. In early 2025, DeepSeek tore open the text-model market with open weights plus rock-bottom prices, cracking the “closed-source API at monopoly pricing” model. MiniMax H3 is now replaying that path in video generation — and then some:
- Open-sourcing the leverage: H3 is the first flagship video model to go open source; weights are public and enterprises can deploy it on-premise.
- Price-killer economics: official pricing is already a third of mainstream models, and third-party platforms have cut beneath that.
- Ecosystem distribution: platforms like Metaso AI have dropped deployment friction to zero — open a web page and generate, no CUDA or GPUs required.
But H3 goes one step beyond DeepSeek: it packed post-production into the model itself. Previous AI video models delivered raw footage — no voiceover, no subtitles, no packaging; the rest happened manually in After Effects, Premiere, or CapCut. H3 understands creative intent, generates the subject footage, matches sound effects and music, renders subtitle text, and completes motion-graphics packaging — one generation pass delivers a finished cut.
Concretely: an e-commerce ad used to require a full chain — text-to-video for product footage, TTS for narration, subtitle pass in CapCut, motion effects in AE. With H3, feed it a product photo and an ad script, and the single output already includes shelf-panning camera moves, narration, subtitles, bouncing price tags, and lighting atmosphere — every visual element aligned with the scene’s space, light, and rhythm.
Adobe’s fiscal 2025 numbers put Creative Cloud global revenue at roughly $14.5 billion, with video post-production tools taking a substantial share. If H3’s AI-native post-production keeps delivering, 10–30% of traditional post workflows are candidates for replacement.
Winners and the Anxious
- Creators — the biggest winners. When per-generation cost drops from “a meal” to “a candy,” the logic of experimentation inverts. A rejected camera angle used to mean an abandoned idea; at 0.09 RMB per second you generate five versions of one concept and pick the best. Creative momentum no longer dies under wallet pressure.
- Enterprise buyers — open weights plus on-premise deployment opens a market closed-source models had locked out. Film studios will not feed brand characters and unreleased footage into public APIs; advertisers cannot tolerate brand visuals drifting randomly per generation; game studios will not expose core IP to third-party models. Data-sovereignty-sensitive, high-value customers finally have a viable video-generation path.
- Closed-source vendors — under the most pressure. For the past six months, video generation was dominated by a few closed models, where price hikes, queues, waitlists, and account reselling became routine. H3’s open-plus-cheap strategy turns the competition from “who charges more” into “who performs better.” Once users stop being locked to a single supplier, pricing power starts migrating.
- MiniMax itself — H3 is a deliberate strike after a turbulent stock price since its January listing. One model doesn’t reshuffle the table, but it proves MiniMax can still flip a game from scratch.
Why This Is the Moment to Enter
For creators, developers, and enterprise decision-makers, the window is now:
- Creators: open Metaso AI or similar platforms and use H3’s online version directly — no GPU, no deployment, prompt in and 2K video out. The cost is low enough to brainstorm with; run every creative concept you have.
- Developers: H3’s weights are open and the API is compatible with the official interface. Integrate it into ComfyUI workflows or deploy it privately; Intel has shipped Day 0 support and the Arc Pro B70 runs it locally.
- Enterprises: if you hold brand IP, unreleased assets, or sensitive data, H3’s on-premise option is the first genuinely viable video-generation path. Video generation is no longer a single-choice question of handing your data to someone else.
For the past six months, the video-generation track was defined by arrogance. H3 has put flagship quality, 2K resolution, open weights, and low prices on the table in one move. This is not a model upgrade — it is an inflection point for the industry’s entire business model.
