Hailuo AI: The Physics Are the Point

Hailuo is the model I reach for when something in the shot has to obey gravity. That sounds like a narrow reason to keep a subscription, and it is — but "the object fell the way an object falls" turns out to be the difference between a clip people watch and a clip people scroll past.

This is what MiniMax's generator is genuinely best at, what it costs once you understand the credit maths, and the specific jobs I hand to it instead of the models I use for everything else.

What Hailuo is, and why the physics claim isn't marketing

Hailuo AI is MiniMax's video generator. The line this guide is built on is Hailuo 2.3, which picked up the "physics champion" label on WorldModelBench for mass conservation, fluid dynamics and spatial-temporal consistency — the three things that make generated footage read as fake when they go wrong.

Those benchmarks usually mean very little. This one matches what I see. Pour liquid, drop something, let fabric swing, and Hailuo keeps volume and momentum consistent across the clip where other models quietly cheat — water that appears from nowhere, a coat that stops moving a beat before the body does. If your shot contains motion with real weight behind it, that consistency is the whole game.

What changed with H3 (Hailuo 3.0), 31 July 2026

MiniMax shipped H3, also called Hailuo 3.0, on 31 July 2026, and it is a bigger jump than the 2.x point releases. H3 is omni-modal: it takes text, images, video and audio as input and returns up to 15 seconds of 2K video at 24fps with native stereo sound — dialogue, effects and ambience generated in the same pass as the picture rather than layered on afterwards. It also ships with open weights, which none of the closed frontier models do.

On price, the published rate is $0.13 per second of 2K (about $7.80 a minute). A cheaper 768p tier at $0.09/s was announced but is still in closed beta at the time of writing, so plan around the 2K number. Reference audio is free and the first few reference images are free before a small per-image charge.

An honest caveat: everything above is confirmed from MiniMax's own release and the independent write-ups, not from my own runs yet. The rest of this guide is written from real generations on 2.3, and I would rather flag the difference than blur it. What I am testing first is whether the physics advantage below survives the move to native audio — if it does, the interesting question becomes whether H3 replaces two tools at once. I will update this section once the clips exist.

The model line, decoded

The naming is genuinely confusing, so here is what the versions actually mean in practice.

Hailuo 2.3 is the current flagship and the one the physics results belong to. Use it when motion realism is the point of the shot.

Hailuo 2.3 Fast cuts cost by up to half and is built for image-to-video iteration. This is where I spend most of my credits — you iterate on Fast until the shot is right, then re-render the keeper on the full model.

Hailuo 02 is the cinematic architecture built on MiniMax's NCR framework: roughly three times the parameters of the previous generation, four times the training data and about 2.5x faster inference. It is the more filmic look.

Hailuo 01 variants remain available on higher tiers. There is rarely a reason to choose them now.

What it costs

Hailuo runs on credits rather than clips, which makes headline pricing misleading until you convert it.

Convert that and the picture gets interesting: roughly $0.30 for a 10-second 1080p clip, against about $1.20 for Runway and around $3.50 for Sora. That is the number that keeps Hailuo in my stack — not because cheap matters on its own, but because cheap changes how many attempts you can afford, and attempts are what produce good clips.

The free tier is also unusually honest. Two to three videos a day is enough to evaluate the model properly before paying, which is more than most competitors allow. If you're assembling a stack on nothing, it belongs on the shortlist in the free generators roundup.

What I actually use it for

I don't run everything through Hailuo. I run specific shots through it.

Anything with liquid, cloth or falling objects. This is the obvious one and it is worth the switching cost on its own.

Image-to-video on a product shot. Start from a still you already control, let Hailuo add motion, and you keep the product accurate while getting movement that doesn't warp it. The general approach is the same one in my image-to-video workflow, and Hailuo's physics handling makes it the safest model for it.

Cheap iteration before an expensive render. Fast variant, ten attempts, pick one. Then decide whether the keeper is worth re-rendering at full quality — often it isn't, and 1080p on Fast is fine.

The honest limits

Three things I would want told to me before subscribing.

Character consistency is not its strength. Hailuo is a physics model, not an identity model. Across separate generations the same person drifts. If your project needs one recognisable character across many shots, solve that with the approaches in keeping characters consistent, or pick a model built around reference elements.

Credits obscure real cost. A 15-credit clip and an 80-credit clip look similar in the interface and differ by more than five times in price. Work out your cost per clip in currency before you plan a batch, or the month gets expensive quietly.

Prompt adherence is good, not exact. It reads physical description well and camera direction less precisely than Kling. If your shot is defined by a specific camera move rather than by what happens in frame, you may fight it.

How I'd evaluate it in an evening

Use the free tier and spend your two or three daily clips on the hardest thing you have — not a talking head, not a slow pan. Pour something. Drop something. Have fabric move. That is the axis Hailuo is optimised for, and testing it on an easy shot tells you nothing you couldn't learn from a cheaper model.

If those clips hold up, the $7.99 tier is among the best value in AI video right now, and the Fast variant makes iteration genuinely affordable. If your work is mostly people talking to camera with consistent faces, spend the money elsewhere — you would be paying for a strength you never use.

Come compare notes

Looking for something else? Browse all 72 AI video guides in one list.

Hailuo is the one where the tier you pick tells you almost nothing about how much video you get, because it bills in credits and the same credit buys wildly different footage depending on resolution and duration. There is also a mechanic that quietly pushes most people one tier higher than their arithmetic suggested: credits reset monthly and do not roll over, so the plan you need is the one that covers your heaviest week, not your average one. The credit maths in the Hailuo AI pricing breakdown.

Already have the narration written? Generating straight from a script is a shorter path than prompting shot by shot.

Hailuo splits opinion more than most models, and I think it's because people evaluate it on the wrong shots. In my free community we swap actual prompts, credit costs per usable clip and side-by-side tests across Hailuo, Veo, Kling, Grok and Seedance — including the generations that failed, which is where the useful information lives. Join the free AI Video Generator community on Skool and post the shot you can't get right; someone has almost certainly tested it.