MiniMax H3 review (Hailuo 03): native audio, 2K output, and what a run really costs
Published 2026-08-12 · Data measured 2026-08-12
MiniMax H3, also shipped as Hailuo 03, cost $0.76 for a 10-second 768p text-to-video run on Route A on 2026-08-12. We measured it before adding the model to our multi-model video service, where every generation produces a charge on our account. The more useful finding was its reference metering. Across runs on 2026-08-10 and 2026-08-12, reference video seconds added to billed time, while a two-image 10-second 2K run on 2026-08-12 cost the same as text-to-video.
TL;DR
- Route A charged $0.76 for 10 seconds at 768p, or $0.076 per output second, on 2026-08-12.
- With the same 12.67s reference video, a 5-second output cost $1.368 and a 10-second output cost $1.748 across 2026-08-10 and 2026-08-12. The unit-rate check differed by 0.4% when billed time included output plus reference video.
- Requested 2K produced 2560×1440, while requested 768p produced 1344×768, verified with ffprobe on 2026-08-12.
- Choose Seedance 2.5 instead when you need a 30-second output, a 30/10/10 reference set, or a 480p tier. MiniMax H3 topped out at 15 seconds and 9/3/3 references, and its lowest tier is 768p with no 480p option, when checked on 2026-08-12.
Method snapshot
On 2026-08-08, we recorded three wall-clock runs for 5-second 768p text-to-video at 137 / 150 / 243 seconds, and confirmed that an audio file could be the only reference. On 2026-08-12, we recorded one run each for 10-second 2K text-to-video, 10-second 768p text-to-video, and a two-image 10-second 2K reference job. The reference-metering test used the same 12.67s clip with 5-second and 10-second outputs across 2026-08-10 and 2026-08-12. We did not use one prompt across every configuration, so this is not a controlled visual-quality comparison.
Measured results: cost and latency
Route A cost per run
These are Route A account charges self-reported in USD and verified on 2026-08-12. Normalized rates keep the original resolution visible.
| Configuration | Account charge | Normalized result |
|---|---|---|
| t2v, 10s, 768p | $0.76 | $0.076 per output second |
| t2v, 10s, 2K | $1.235 | $0.1235 per output second; 2K factor 1.625 in the two-point check |
| r2v, 10s, 2K, two image references and no reference video | $1.235 | Same as the 10s 2K t2v run |
| r2v, 5s, 768p, plus a 12.67s reference video | $1.368 | $0.0774 per billed second |
| r2v, 10s, 768p, plus the same 12.67s reference video | $1.748 | Unit-rate difference 0.4% across the metering pair |
The 10-second 2K image-reference run and the 10-second 2K text-to-video run both charged $1.235 on 2026-08-12. That is the direct receipt evidence behind the image-reference result. It is not an estimate from a public price sheet.
Wall-clock latency by run
We preserve every measured run instead of reporting only an average.
| Configuration | Wall clock | Measurement date |
|---|---|---|
| t2v, 5s, 768p | 137 / 150 / 243 seconds | 2026-08-08 |
| t2v, 10s, 2K | 359 seconds | 2026-08-12 |
| t2v, 10s, 768p | 264 seconds | 2026-08-12 |
| r2v, 10s, 2K, two image references | 497 seconds | 2026-08-12 |
The 2026-08-08 latency row has three runs. Each 2026-08-12 latency row has one run. The data shows observed spread and individual outcomes, not a service-level guarantee.
Verifiable specifications and the Seedance 2.5 boundary
Specifications were checked on 2026-08-12. The Seedance 2.5 comparison uses the previously measured 2026-08-08 data rather than a new run.
| Decision variable | MiniMax H3 | Seedance 2.5 |
|---|---|---|
| Output length | 5 to 15 seconds, integer values | Up to 30 seconds |
| Resolution choices | 768p or 2K | Includes a 480p tier |
| Reference capacity | 9 images, 3 videos, 3 audio files | 30 images, 10 videos, 10 audio files |
| Measured normalized cost | $0.076/s at 768p on 2026-08-12 | $0.0968/s at 480p on 2026-08-08 |
The cost row is not a same-resolution comparison. It describes the measured entry tier for each model, with the resolution left visible so the numbers are not mistaken for a controlled price benchmark.
MiniMax H3 also generated native synchronized audio. Two inspected outputs carried an AAC audio track when checked with ffprobe on 2026-08-12. An audio file also succeeded as the only reference in a create request on 2026-08-08.
For API integration, the resolution parameter is quality, with 768p and 2k as the accepted values. Sending resolution returned HTTP 400 on 2026-08-12. Reference audio belongs in audio_urls; sending reference_audio_urls returned HTTP 400 on 2026-08-12.
Exclusive experiment: what reference mode actually meters
The key question was whether Route A billed only output time or also billed reference inputs. We used the same 12.67s reference video, then requested a 5-second output and a 10-second output across 2026-08-10 and 2026-08-12.
| Output request | Reference video | Charge | Metering check |
|---|---|---|---|
| 5s at 768p | 12.67s | $1.368 | $0.0774 per billed second |
| 10s at 768p | 12.67s | $1.748 | Unit rate differed by 0.4% |
The two charges align when billed seconds are output seconds plus reference-video seconds. The separate two-image 10-second 2K job charged $1.235 on 2026-08-12, exactly matching the 10-second 2K text-to-video charge on the same date.
Reference video time is metered; image references did not add billed time in this Route A test. The supplied routing record says the same output-only rule applies when audio is the sole reference, but it does not include a separate USD receipt value for that audio-only job, so we do not publish an audio-specific price until that receipt exists.
Resolution and audio were checked from the output file
Route A returned 2560×1440 after a 2K request and 1344×768 after a 768p request on 2026-08-12. The requested resolution was present in the delivered media rather than only in request metadata. Two inspected outputs also contained an AAC stream on 2026-08-12.
This command reproduces the output inspection without relying on the route response:
ffprobe -v error -show_entries stream=codec_type,codec_name,width,height -of json output.mp4
The production MiniMax H3 page also hosts our honey macro, rain and synchronized audio, and dual-image cello reference samples measured for this release.
The outputs themselves, attached
Cost tables do not show what the model produces. These are the three release samples measured for this article, self-hosted, with the checkable claims each one carries.
Native 2K, verified on the delivered file
This 10-second macro run was requested at 2K and the delivered file measures 2560×1440 by ffprobe (2026-08-12), the same check described in the resolution section. The claim this clip carries is simple: the resolution you pay for is present in the file, and the level of micro-detail (bubbles, sugar crystals, crumb pores) is what that resolution buys.
Synchronized audio, published for your own ears
The prompt scripted three timed sound events: rain on glass throughout, a tram pass swelling near 4s, and a cup set down near 7s. The delivered file carries an AAC audio track (verified 2026-08-12). We are not going to grade our own audio sync; the clip is here so you can judge whether the scripted events land where the picture says they should.
Two image references composed into one scene
Inputs: a cellist character still and an empty flooded-hall location plate. The output had to place the one inside the other and keep both consistent.
In our review of this run: the dress, hairstyle and instrument follow the character still; the column geometry, skylight beam and water reflection follow the location plate; the reflection tracks her motion. The billing side of this exact run is in the reference-metering table above: two image references, 10 seconds at 2K, charged the same as text-to-video. One run, published as evidence rather than a benchmark.
When MiniMax H3 is not the right choice
Use Seedance 2.5 for a 30-second shot. MiniMax H3 stopped at 15 seconds in the specification checked on 2026-08-12.
Use Seedance 2.5 when the brief needs more than 9 images, 3 videos, or 3 audio files. Its reference envelope reaches 30 images, 10 videos, and 10 audio files, while MiniMax H3 uses the smaller 9/3/3 envelope checked on 2026-08-12.
Use Seedance 2.5 when a 480p low-cost tier matters. MiniMax H3 started at 768p when checked on 2026-08-12.
The available price observations also use different resolutions: MiniMax H3 measured $0.076/s at 768p on 2026-08-12, while the existing Seedance 2.5 measurement is $0.0968/s at 480p on 2026-08-08. Do not treat those as a same-resolution quality or price comparison.
Caveats
- The 5-second 768p latency row has three runs from 2026-08-08. The other latency rows have one run each from 2026-08-12.
- The metering decision uses two output lengths with the same 12.67s reference video, measured across 2026-08-10 and 2026-08-12. It does not establish every possible reference combination.
- The 2K pricing factor of 1.625 is a two-point check from 2026-08-12, not a broad price curve.
- Provider prices can change. Every USD figure in this article is a Route A account charge verified on 2026-08-12.
- We did not run a controlled visual-quality comparison. A failure-count sample was not collected, so no failure-rate percentage is reported.
Method
Cost values came from Route A supplier receipts self-reported in USD and verified on 2026-08-12. We retained the charged amount for each run and normalized it by the billed seconds specified in the handoff. Route labels stay anonymous because the procurement channel is not part of the model result.
Latency is wall-clock time for each recorded run. We report the 2026-08-08 three-run sequence and each 2026-08-12 single run separately. We did not turn them into an average.
For the reference-metering test, the same 12.67s reference video was used with 5-second and 10-second outputs across 2026-08-10 and 2026-08-12. For the no-reference-video control, the 10-second 2K two-image job was compared with the 10-second 2K text-to-video job on 2026-08-12. Both charged $1.235.
Resolution and audio were checked on the delivered files with ffprobe. The observed dimensions were 2560×1440 for a 2K request and 1344×768 for a 768p request on 2026-08-12. Two inspected outputs contained AAC audio.
Inspect the release samples and settings on MiniMax H3. Reproduce the multi-reference workflow in Reference to Video.