H3 Max vs MiniMax H3: Same Family, and the Price Gap Has an Expiry Date

MiniMax H3 Max and MiniMax H3 share a family tree, which is exactly what makes the comparison strange: the post-trained derivative outranks its own parent on both Artificial Analysis with-audio leaderboards, generates faster than real time, and is cheaper per minute today — and yet the base MiniMax H3 still holds the two assets the post-train gives up, 2K output and open weights, while the price gap between them evaporates on September 1. MiniMax H3 Max sits at #1 in image-to-video (Elo 1,204) and #3 in text-to-video (Elo 1,235) with audio; the base MiniMax H3 sits at #3 (1,184) and #4 (1,226). Same family, same native-audio output, and in the same week the answer to two genuinely different questions.

One family, two serving decisions

MiniMax H3 arrived on July 31, 2026 as MiniMax's first open multimodal video model — the family is also branded Hailuo 3.0 — and its weights went public on August 3. MiniMax H3 Max is the post-train fal built from that open release: tuned for stronger prompt adherence and better aesthetics, then co-optimized with fal's inference stack for higher throughput. Both models generate clips with native audio and share the H3 output envelope, but they diverge at exactly the decision point: fal took the open release and made it faster, higher-scoring, and cheaper-per-minute right now, while MiniMax kept the full 2K pipeline and the downloadable weights behind a licensed API.

The scoreboard gap

On the with-audio boards, captured 2026-08-27, the post-train beats the base on both: 1,204 to 1,184 in image-to-video, and 1,235 to 1,226 in text-to-video. The gaps are modest — 20 and 9 points — but they are consistent, and the text-to-video podium is a six-point photo finish (Wan 3.0 at 1,240, Gemini Omni Flash at 1,237, MiniMax H3 Max at 1,235) where the post-train's margin over the base is almost exactly the width of the whole top of the board. These are blind human-preference Elo scores from Artificial Analysis' arenas, not vendor numbers, and both boards drift as new votes arrive.

• Image-to-video (with audio) — MiniMax H3 Max #1, Elo 1,204. MiniMax H3 #3, Elo 1,184.

• Text-to-video (with audio) — MiniMax H3 Max #3, Elo 1,235. MiniMax H3 #4, Elo 1,226.

Specs side by side

• Price per minute — MiniMax H3 Max: $2.40 at 768p on a 50% launch discount, $4.80 standard after September 1. MiniMax H3: $4.80 at 768p, $7.80 at 2K.

• Max resolution — MiniMax H3 Max: 768p. MiniMax H3: 768p on the open weights, 2K on the hosted API.

• Clip length — MiniMax H3 Max: 5-15s. MiniMax H3: 4-15s. Both generate native stereo audio.

• Generation speed — MiniMax H3 Max: ~0.55x real-time on fal's sample timing (5s clip in 2.77s, vendor-reported). MiniMax H3: no public figure.

• Weights — MiniMax H3 Max: intent to release, not shipped. MiniMax H3: open August 3, community license.

• Where to call it — MiniMax H3 Max: the developer's own API only today. MiniMax H3: MiniMax's API and routed platforms including OrcaRouter.

The price gap has a deadline

The cleanest number in this matchup — $2.40 versus $7.80 per minute on Artificial Analysis — is partly a promo artifact, and it is worth being honest about which parts. $2.40 is MiniMax H3 Max's discounted 768p rate, live until September 1. $7.80 is the base MiniMax H3's 2K rate, which the post-train cannot reach at all. The apples-to-apples comparison is at 768p: there the base MiniMax H3 lists at $0.08 per second, or $4.80 per minute, and the post-train's standard rate after the discount is exactly that — $4.80. So today's price edge is real but temporary, and its size depends on you comparing like with like. At high volume the difference while the promo lasts is still decisive: a thousand minutes is $2,400 on the post-train against $4,800 on the base at the same resolution, before discard rate — and both models bill rejected clips.

Speed changes what you can build

The second differentiator is operational rather than numeric. fal's model page shows a 5-second clip rendering in 2.77 seconds of inference — a ~0.55x real-time factor, vendor-reported and not yet independently reproduced. Sub-real-time generation changes the shape of the workflow: instead of queuing a clip and waiting, you can generate, judge, and regenerate inside a loop, which matters for interactive previews, agent-driven video, and high-volume A/B testing of prompts. The base MiniMax H3 is no slouch on throughput either, but it does not publish a comparable figure, so the speed advantage is currently one-sided on paper. Just keep the resolution asterisk attached: H3 Max's speed is partly a function of its 768p ceiling, and the same loop at 2K is not on offer.

The 2K question

MiniMax H3 Max cannot go above 768p, and neither can the base model's open weights — the 2K regeneration step is an API-only module on the base's hosted endpoint. If your deliverable is 1080p or above, the decision is made for you: that is the base MiniMax H3's 2K API, at $0.13 per second ($7.80 per minute). If your target is 768p — social cutdowns, ad variants, e-commerce loops, quick reference shots — the resolution gap is irrelevant and the post-train's price and speed win the argument while the promo lasts. "Max" in the name refers to the leaderboard and the throughput, not the resolution ladder, and buying it means buying the ceiling it ships with.

Weights: open vs. promised

The openness gap is where the base model's edge is structural, not temporal. MiniMax H3's weights have been downloadable since August 3, under a custom community license that per MiniMax's own terms excludes use of the model and its outputs in the EU, the UK, South Korea, and the US, and requires prior written authorization for commercial use above $20 million in annual revenue. MiniMax H3 Max's weights are a stated intent from fal, not a shipped artifact — and if they land under anything similar, "open" carries the same territory drawn around it, with the added wrinkle that a third-party post-train of a licensed open release sits in a greyer legal spot than the original. For a team that needs to self-host or that operates in an excluded territory, the base model is the only current option, license caveats and all.

Which one to call

The choice today comes down to which constraint binds first. Need the current #1 image-to-video model, want sub-real-time iteration, and can live with 768p? MiniMax H3 Max is the value pick while the discount runs — just put the September 1 price change in your cost model. Need 2K, downloadable weights, or a single API key for the whole H3 family? That is the base MiniMax H3, and it is the half of the family already on OrcaRouter at the provider's list price — $0.08 per second at 768p, $0.13 at 2K — with 0% markup and automatic failover, so you are not locked into one serving path while the H3 Max routing picture settles. When the post-train reaches a routed provider it will appear at list price the same day, and the swap becomes a config change rather than a rewrite.

The bottom line

This is the rare case where the derivative is the better value — for a while, and within a resolution. MiniMax H3 Max outranks MiniMax H3 on both with-audio boards, generates faster than the clip plays, and undercuts it at the resolution both models share, but the price edge expires September 1, the resolution stops at 768p, and the weights are a promise. MiniMax H3 keeps 2K, open weights, and router integration today. Buy the post-train for speed and the current leaderboard crown; buy the base for the ceiling, the license, and the fact that you can call it through one API right now.