HiDream-O1-Video-1.0 vs MiniMax H3: A Six-Point Gap and a Licence That Excludes Four Countries

Six points. That is the entire separation between HiDream-O1-Video-1.0 and MiniMax H3 on the only board where both appear: 1,175 ±10 for HiDream against 1,181 ±8 for H3 on Artificial Analysis's image-to-video leaderboard, read 10 October 2026. The intervals run 1,165–1,185 and 1,173–1,189. They overlap almost completely. Any page that opens by declaring a winner on that board is reading noise.

The tie is real and it is not the story. The story is what surrounds it: three boards against one, a price difference that depends entirely on which of H3's two hourly-cost readings you take, and the fact that one of these models can be downloaded — under terms that bar use in the United States, the United Kingdom, the European Union and South Korea unless you apply for permission. The score is the least decisive number in this comparison.

Both boards, both models, and where the coverage stops

AA-Video-I2V v1.0, audio kept in:

• MiniMax H3 Max — 1st, Elo 1,195 ±9, 5,894 votes, August 2026, $4.80/min

• MiniMax H3 — 2nd, Elo 1,181 ±8, 7,284 votes, July 2026, $7.80/min

• HiDream-O1-Video-1.0 — 6th, Elo 1,175 ±10, 3,231 votes, August 2026, $5.80/min

AA-Video-T2V v2.0, a separate board with its own scale and its own vote pool:

• MiniMax H3 (768p) — 4th, Elo 1,137 ±9, 8,315 votes, July 2026, $4.80/min

• MiniMax H3 Max — 5th, Elo 1,130 ±10, 5,719 votes, August 2026, $4.80/min

• HiDream-O1-Video-1.0 — not listed

AA-Video-Editing v1.1:

• MiniMax H3 (768p) — 4th, Elo 1,119 ±6, 7,023 votes, July 2026, $9.60/min

• HiDream-O1-Video-1.0 — not listed

MiniMax's H3 family is three rows across three boards with three different per-minute readings, because editing and text-to-video and image-to-video are metered differently on the same weights. HiDream's model is one row on one board. The gap in the score is six points and inside the noise; the gap in coverage is two entire task categories wide.

There is a second reading of the I2V table worth naming, because it cuts against the obvious conclusion. HiDream's 1,175 rests on 3,231 votes. MiniMax H3's 1,181 rests on 7,284 — more than twice as many — and H3 Max's 1,195 on 5,894. On a board where the whole top six spans twenty points, the model with the smallest vote pool has the widest interval and the most room to move. Six points of separation between a 3,231-vote measurement and a 7,284-vote measurement is not evidence of parity or of inferiority; it is evidence that the question has not been settled.

The licence, before anything else

If you are in one of the excluded territories, the rest of this page is academic, so it goes first.

MiniMax H3's base weights are downloadable under MiniMax's community licence — not Apache, not MIT — and that licence excludes the United States, the United Kingdom, the European Union and South Korea, where use requires an application to MiniMax. Commercial use above $20 million in annual revenue requires prior written authorisation. What is open is the base checkpoint specifically: the Context-IR prompt-expansion module and the 2K regeneration path stay API-side and were never part of the release.

HiDream-O1-Video-1.0 publishes no weights at all on Artificial Analysis's record, and no first-party repository is reachable for the video model. That is a different kind of closed: MiniMax shipped something with terms you might be able to meet, and HiDream shipped nothing to meet.

The practical consequence is narrower than "open versus closed" makes it sound. If you are outside the four excluded territories and self-hosting is on your roadmap, H3's weights are a real option and HiDream's are not. If you are inside them, H3 is an API product on the same terms as HiDream's, and the licence has cost you nothing but a paragraph of reading.

Specifications, and how much of each is checkable

• Length — MiniMax H3: 4 to 15 seconds, integer values only, provenance MiniMax's own platform specification. HiDream-O1-Video-1.0: the board records it as a released production model; its served contract documents a much narrower request than its announcement describes, and no first-party page reachable today publishes a duration ladder.

• Resolution — MiniMax H3: 768p natively, with 2K available through a regeneration module that is API-only and therefore unavailable to anyone self-hosting the weights. HiDream: priced by the board at a 1080p configuration; no vendor resolution list published.

• Frame rate and audio — MiniMax H3 publishes a clear audio specification: 32 kHz stereo generated in the same pass as the video, at 24fps. HiDream's announcement describes natively synchronised audio and HiDream-O1-Video-1.0 appears on Artificial Analysis's with-audio board rather than the silent one, which is independent corroboration that the audio is real. Neither vendor publishes an audio spec of H3's precision for the HiDream model.

• Inputs — MiniMax H3: text, image, video and audio references read as one unified context, up to 9 images, 3 video clips and 3 audio clips. HiDream: its announcement lists text, images and video, but the public request schema documents one reference image plus an optional prompt, which is the shape you would actually integrate against.

• Output control — H3 offers first-frame, last-frame and first-and-last-frame control plus reference-to-video. HiDream's documented public surface offers none of those controls.

The pattern in that list is worth stating plainly. Every line on the MiniMax side comes from MiniMax. Most lines on the HiDream side come from the board or from the announcement rather than from a specification, and where the announcement and the served contract disagree, the served contract is the one that will reject your request.

What each second costs, at the resolution you would buy

MiniMax bills per second of generated output: $0.08 per second at 768p and $0.13 per second at 2K on its own platform. Reference audio input is free; reference images are free for the first five. Those rates render as $4.80 per minute for the 768p configuration on the text-to-video and editing boards, and as $7.80 per minute for the 1080p-equivalent configuration the image-to-video board records.

HiDream's single circulating figure is $5.80 per minute, at 1080p, on the model creator's API at default settings, as recorded by Artificial Analysis.

Read at face value, that looks like HiDream sitting between H3's two readings. It is worth resisting, for two reasons. The $4.80 and the $7.80 are the same model at different resolutions, so comparing $5.80 against either is comparing a resolution to a resolution rather than a model to a model. And the draft-then-upscale trick does not rescue the arithmetic: generating H3 at 768p and regenerating to 2K costs $0.08 plus $0.05, which is exactly the $0.13 per-second 2K rate. It saves money only on takes you were going to discard anyway.

The metric that survives all of this is cost per accepted clip, which folds in retries. On reference-heavy work — recurring characters, a product that has to stay identical across shots — H3's nine image references and three audio references are aimed at exactly the failure mode that generates retakes, and a model that holds the subject in one attempt is cheaper than its meter suggests. On prompt-only work those slots are shelves you will never fill.

Availability, where the asymmetry is unusually concrete

This is the rare matchup where the open-weight model is also the more available one.

MiniMax H3 is in the OrcaRouter catalogue as minimax/minimax-h3, billed at the provider's list price — $0.08 per second at 768p and $0.13 at 2K, the vendor's own rates — with no markup added on our side, so a MiniMax price change is live on our line the same day rather than at renewal. It sits on the same OpenAI-compatible endpoint as the text models, which means a video call is a model-ID change rather than a second integration, and automatic failover moves the request to a healthy serving path before the response starts instead of returning an error to your application.

HiDream-O1-Video-1.0 is not an OrcaRouter route, and this piece is not claiming it is. It is reachable through the vendor's own surfaces and through third-party workspaces that resell it, with the $5.80 per minute the board records as its published economics.

That asymmetry cuts the other way in one place, and it is worth naming. Routing does not fix a licence. If you are in the US, UK, EU or South Korea and you intended to self-host H3's weights, calling it through a router is not a workaround — it is a different product, and the licence question is still yours to answer.

Choosing between them

• Both, if the job is animating a still with sound and nothing more. The two are a statistical tie on that board, and at that point you are choosing on documentation, price and whether you want the licence option. H3 wins on all three in most cases, and HiDream wins on nothing you can currently verify.

• MiniMax H3 alone, if the input is a script rather than a frame. HiDream's documented request will not accept the call, and no first-party text-to-video route is published for it.

• MiniMax H3 alone, if you cut, extend or restyle footage. H3 ranks fourth on the editing board at 1,119 ±6 over 7,023 votes. HiDream is not entered.

• MiniMax H3 alone, if self-hosting is on the roadmap and you are outside the four excluded territories. Nothing on the HiDream side substitutes.

• HiDream-O1-Video-1.0, if you specifically want a second independent vote count behind a newer model and are willing to integrate against a narrow, still-moving request contract. Six points below H3 on that board, at a lower recorded rate, with the caveat that the measurement is the younger and better-established one's junior by a factor of two in sample size.

What would settle it is a HiDream row on a second board. If the model acquires a text-to-video or editing entry, the coverage half of this comparison collapses and it becomes a straightforward two-model price-and-score question, which the current numbers leave open. If it does not, the defensible summary is the narrow one: on the single board both models are entered on, they are tied inside their intervals, and one of them does two more jobs.