Seedance 2.5 vs MiniMax H3 vs FLUX 3: what's verified, and what isn't
Updated July 31, 2026
Three flagship video models shipped in nine days — FLUX 3 on July 23, Seedance 2.5 and MiniMax H3 on the same morning, July 31. The internet is already full of confident rankings of models that essentially nobody has tested side by side. This page separates what is checkable from what is being repeated.
The honest headline: nobody has tested these against each other
We looked, because this page was supposed to be a shootout. It isn't one, and here is why:
- No arena scores. We checked the Artificial Analysis video arena on July 31: none of the three is on it. The board still ranks the previous generation.
- No independent head-to-head.The comparison articles published this week are built from vendor specs and each other. The few that admit this say the same thing we do — that a fair test isn't possible until access is.
- Access is the blocker. One of the three is buyable, one is behind a Chinese consumer app, one is behind an application form. There is no configuration in which one person can run all three on the same prompt today.
What is verified, side by side
Every cell below comes from a vendor's own launch material or a check we ran ourselves on July 31, 2026.
| Seedance 2.5 | MiniMax H3 | FLUX 3 | |
|---|---|---|---|
| Launched | Jul 31, 2026 | Jul 31, 2026 | Jul 23, 2026 |
| Can you use it? | Jimeng AI / Doubao Pro only — no API | Yes — fal.ai, Hailuo app, MiniMax API | Application-gated early access |
| Max single generation | 30s, extendable into minutes | 15s | 20s |
| Resolution | Up to 4K (2.0 lineage) | Native 2K, 24fps | 720p in EA, 1080p expected |
| Reference budget | 30 images + 10 videos + 10 audio | 9 images + 3 videos + 3 audio | Up to 10 images, plus video and audio |
| Audio | Native, improved over 2.0 | Native stereo | Native, multilingual dialogue |
| Editing | Timestamp-addressed; green screen, viewpoint, camera move | Instruction-based editing, V2V motion transfer | Video-to-video, keyframes, continuation |
| Also generates | Video only | Images, audio | Images, audio, robot actions |
| Open weights | No | Promised “within days” | FLUX 3 Dev, undated |
| Price you can buy at | None published | From $0.14/s at 2K | None published |
Full working on each: Seedance 2.5, MiniMax H3, FLUX 3.
Launch-week claims that don't survive a check
These are all circulating right now, some of them on pages that rank well. Each one is wrong against the vendor's own material or against a check anyone can repeat.
| The claim | What we found | |
|---|---|---|
| “H3 does 4K at 60fps, up to 30s” | Seen on at least one guide site | MiniMax's own launch post says up to 15s at 2K; fal and Atlas both sell it as 2K/24fps |
| “H3 was unveiled at WAIC on July 17” | Repeated in several write-ups | MiniMax published the H3 launch post on July 31, and its own site tagged the model NEW that day |
| “H3 is gated, no public API endpoint” | Stated as current fact | fal.ai had three public endpoints live on launch day (minimax/h3/text-to-video, /image-to-video, /reference-to-video) |
| “Seedance 2.5's API opens August 7” | One site states it as a date | No other source carries it and ByteDance says only “soon”. The same rumour chain produced July 6, 13 and 16 — all of which failed |
| “H3 costs $0.13/s at 2K” | Widely quoted, including by us that morning | fal doubled it to $0.26/s within twelve hours of launch. Any price without a date on it is unreliable this week |
Why so much of this is wrong
A launch week is when content farms harvest the most traffic and the fewest people can check anything. Two of the three models can't be run by the authors writing about them, so specs get copied between posts and mutate on the way. The tell is a page quoting a firm number with no date and no source link.
What the vendors claim about themselves
Worth reading, worth discounting. Black Forest Labs ran a preference study putting FLUX 3 ahead of Luma Ray 3.2 (93%) and Runway Gen-4.5 (77%) — but at 52% against Seedance 2.0, a tie with the generation Seedance 2.5 replaces, and BFL labels the chart a preliminary evaluation of a pre-release checkpoint. MiniMax claims H3 runs at under a third of mainstream per-second pricing at 2K, which holds at Atlas Cloud's $0.14 and doesn't at fal's $0.26. ByteDance published capabilities and demo prompts for Seedance 2.5 but no comparative numbers at all.
The test we will run
Publishing the protocol before we have access is the point — it stops us from picking a prompt afterwards that flatters a conclusion. When all three can be run, each gets the same brief at matched duration and resolution:
- One action with a physical outcome — an object leaves a surface, falls, and breaks or lands. Scored on whether the event completes and whether the fragments behave plausibly, not on whether the frame is pretty.
- One camera move that has to survive the action — scored on whether the geometry of the room stays consistent through it.
- One sound event— scored on the frame the sound starts relative to the impact. This is where jointly generated audio either proves itself or doesn't.
- One identity to holdacross the full duration, using each model's own reference mechanism.
- Measured cost per finished clip, including retries — the number that decides production use, and the one nobody publishes.
Outputs go up unedited, failures included, with the prompts and the bill. Our prompt sets for Seedance 2.5 and FLUX 3 are already written for this.
What to do this week
- Need output now:MiniMax H3 is the only one of the three you can buy, and Atlas Cloud is currently about half fal's 2K rate for the same model.
- Need 30-second takes or 50 references:wait for Seedance 2.5's API. Nothing else does either.
- Building on open weights:H3's are promised within days and FLUX 3 Dev's are undated. Neither exists yet — we check the Hugging Face orgs daily.
Frequently asked questions
- Which of the three is best?
- Nobody can answer that honestly today, and you should distrust any page that does. As of July 31, 2026 none of the three appears on the Artificial Analysis video arena — we checked the board directly — and no independent side-by-side test has been published. Every quality claim in circulation traces back to a vendor's own evaluation or to another blog repeating one.
- Which one can I actually use right now?
- MiniMax H3, and it isn't close. H3 has three public endpoints on fal.ai plus the Hailuo app, live the day it launched. Seedance 2.5 is inside Jimeng AI and Doubao Pro only — no API, nothing outside China. FLUX 3 requires an application to Black Forest Labs and has no published price at all.
- Which generates the longest clip?
- Seedance 2.5: 30 seconds in a single generation, extendable in rounds into several coherent minutes. FLUX 3 does 20 seconds and chains clips into multi-shot sequences. MiniMax H3 tops out at 15 seconds. Length is one of the few dimensions where the three are cleanly ordered on confirmed numbers.
- What do they cost?
- Only H3 has a price you can act on, and even that moved on launch day: Atlas Cloud lists $0.14 per second at 2K while fal.ai went from $0.13 to $0.26 within twelve hours. Seedance 2.5 is bundled into Chinese consumer apps with no per-second rate, and FLUX 3 has published nothing. Any article quoting a firm price for the other two is inventing it.
- When will there be real benchmarks?
- When two things happen: Seedance 2.5's Volcano Ark API opens, and the arenas add the new models. We check ByteDance's ModelArk catalog, fal, kie and the arena boards daily and will publish our own matched-prompt test within days of access, including the failures.