MiniMax may have a new Hailuo video model coming, but the public record does not support treating H3 as a released open-weights product yet.
The AI video market moves fast enough without pretending a promise is a product. As of August 3, MiniMax's own public API documentation still lists MiniMax Hailuo 2.3 as its current video model, with text-to-video and image-to-video output up to 10 seconds and 1080p at 24 frames per second. It does not list H3. It does not list Hailuo 3.0 as an API model. It does not publish H3 weights, a technical report, or a license file for developers to inspect.
That's the story.
The earlier claim that MiniMax released H3 on July 31 as a 2K, native-audio, open-weights video model could not be confirmed in live search against MiniMax's official site, its API platform, or its Hugging Face organization. There are SEO pages using the Hailuo 3.0 name, but they are not enough to support the article's release claim. If you're building a startup on top of AI video infrastructure, that distinction is not pedantry. It's the line between evaluating a model and building on vapor.
The real MiniMax facts are less dramatic #
MiniMax is still worth watching. Its Hailuo line is a serious video product, and its official API pricing page currently lists Hailuo 2.3 video generation at $0.28 for a 768p six-second clip, $0.56 for a 768p ten-second clip, and $0.49 for a 1080p six-second clip. That is already cheaper than several closed video APIs, depending on the shot length and resolution you actually use.
But cheap is not the same as open.
MiniMax's open-weights story is strongest today around its M-series language and multimodal models, not Hailuo video. Artificial Analysis reported in June that MiniMax-M3 was planned as an open-weights model, and MiniMax's Hugging Face organization now shows public M3 activity. Separate coverage of the M3 license has also pointed out the catch: MiniMax's community license is not MIT or Apache-style permissive licensing. Commercial users need to read the terms before they build anything real.
You should treat any future Hailuo open-weights release the same way. Wait for the repository. Read the license. Check the model card. Then price the infrastructure.
OpenAI and Google are still the benchmarks #
The pricing comparison also needed correction. OpenAI's own developer docs list Sora 2 Pro at $0.30 per second for 720p, $0.50 per second for the 1024p tier, and $0.70 per second for 1080p. The earlier article understated the 1080p number by stopping at $0.50.
Google's Veo numbers are more complicated because the product line keeps shifting across Gemini API, Vertex, and enterprise surfaces. Google's public Gemini API pricing page lists Veo 3 video with audio at $0.40 per second and Veo 3 Fast at $0.15 per second, while newer Veo 3.1 pages describe the model's capabilities rather than giving the clean price table a startup finance team would want. Third-party pricing trackers put Veo 3.1 Standard around $0.40 per second at 720p and 1080p, with 4K higher. Fine, use that as a planning estimate. Don't use it as a sourced fact unless you can tie it back to Google's live pricing page on the day you publish.
Here's the thing: the market pressure is real even if the H3 claim is not. Chinese AI labs have been using price and partial openness as weapons, and American labs have had to respond. DeepSeek did that in reasoning models. Alibaba has done it with Qwen and Wan. ByteDance's Seedance 2.5, announced at Volcano Engine FORCE on June 23 and reported by TechTimes as opening API access through BytePlus on July 16, is pushing on longer video generation with 30-second clips and up to 50 reference inputs.
MiniMax may join that fight with a more open Hailuo release. It probably will. But you don't publish the finish line before the race has been run.
For founders, the practical advice is blunt. Test Hailuo 2.3 if its current API fits your workflow. Compare it with Sora, Veo, Seedance, Wan, and LTX using your own prompts - not leaderboard screenshots. If MiniMax releases H3 weights, inspect the license before you move engineering time toward self-hosting. Don't self-host a model you can't legally deploy at scale. A model you can download but can't legally use at scale is not the same thing as open infrastructure. Also read: AI Agents Are Learning to Lie and Cheat Because Training Rewards Winning • Zhipu AI stays silent as reports point to a trillion-parameter GLM-5.5 • EU AI Act Transparency Rules Take Effect Today Despite the 2027 Delay