# MiniMax H3 (Hailuo) API Pricing: Cost Per Clip at 768P and 2K

> As of November 12, 2026, MiniMax H3 lists at $0.08/s (768P) and $0.13/s (2K); LaoZhang API charges $0.05 and $0.08, so a 5-second 768P clip is $0.25.

- URL: https://blog.laozhang.ai/en/posts/minimax-h3-api-pricing
- Published: 2026-11-12
- Updated: 2026-11-12
- Author: LaoZhang AI Team (https://blog.laozhang.ai/en/about)
- Topic: AI Video Generation
- Tags: MiniMax H3, Hailuo, Video API, API Pricing, AI Video Generation

---
MiniMax H3, the video model also listed as Hailuo 03 on some platforms, is billed by the second of video you ask for. As of November 12, 2026, [MiniMax's pay-as-you-go pricing page](https://platform.minimax.io/docs/guides/pricing-paygo) lists `MiniMax-H3` at $0.08 per second at 768P and $0.13 per second at 2K. LaoZhang API has listed the same model at $0.05 and $0.08 per second since October 24, 2026, which is 37.5% and about 38% below MiniMax's list. A 5-second 768P clip therefore costs $0.25 instead of $0.40, and a 10-second 2K clip $0.80 instead of $1.30.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

To pay those rates, create a token in the `default` group with the Usage first billing mode, send a request with `"model": "MiniMax-H3"` to `https://api.laozhang.ai/hailuo/v2/video_generation`, and read each task's charge in the call logs. You pay for the `duration` you request, from 4 to 15 whole seconds. Reference video adds its own seconds at the same rate, the first five images in a task are free, reference audio is free, and a task that fails is refunded.

Two cases fall outside those numbers. MiniMax's H3 Max, Highspeed and Lite are separate models with their own official prices, and LaoZhang API serves only the full `MiniMax-H3`. If you meant the Hailuo app's credits and monthly plans rather than API time, [Hailuo AI Free Credits: No More Daily Credits, Just a 3-Day Pack](https://blog.laozhang.ai/en/posts/hailuo-ai-free-credits) covers what the app charges per second.

## MiniMax H3 price per second and per clip at 768P and 2K

A MiniMax H3 clip costs its requested length times the rate for its resolution, and nothing else changes the output price. Landscape, portrait and square cost the same, and every clip comes with a stereo soundtrack that has no off switch and no extra fee. Clips usually run 0.1–0.5 seconds longer than the `duration` you set, and on LaoZhang API that extra fraction is free because the bill follows the requested seconds, as the [LaoZhang MiniMax H3 API docs](https://docs.laozhang.ai/en/api-capabilities/minimax-h3-video-generation) state.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

The table multiplies those list prices out for common clip lengths. It is a calculation run on November 12, 2026 from the two price lists, not a set of invoices, and it covers text-to-video output only, with no retries.

| Clip length | 768P on LaoZhang API | 768P on MiniMax API | 2K on LaoZhang API | 2K on MiniMax API |
| --- | ---: | ---: | ---: | ---: |
| 4 seconds | $0.20 | $0.32 | $0.32 | $0.52 |
| 5 seconds | $0.25 | $0.40 | $0.40 | $0.65 |
| 6 seconds | $0.30 | $0.48 | $0.48 | $0.78 |
| 10 seconds | $0.50 | $0.80 | $0.80 | $1.30 |
| 15 seconds | $0.75 | $1.20 | $1.20 | $1.95 |

Turned into a budget, $10 buys forty 5-second 768P clips on LaoZhang API against twenty-five at MiniMax's list price. At 10 seconds the counts are 20 against 12 for 768P, and 12 against 7 for 2K. Because 2K costs 1.6 times as much per second on LaoZhang API ($0.08 vs $0.05) and 1.625 times on MiniMax's API ($0.13 vs $0.08), resolution is the first setting to check when a batch costs more than planned.

## How reference images, video and audio change a MiniMax H3 bill

Only two inputs add money: images beyond the fifth, and reference video. Both price lists give the first five images in a task free; from the sixth, each image costs $0.03 on LaoZhang API and $0.04 on MiniMax's API, and LaoZhang API accepts up to nine images per task. A first-frame or last-frame request holds at most two images and can't be combined with reference images, so animating from frames never adds an image charge. Reference audio is free on both, though LaoZhang API won't accept it as the only reference. A reference video is billed for its own length at the output resolution's per-second rate, on top of the output seconds.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

| Inputs on a 5-second clip | 768P on LaoZhang API | 768P on MiniMax API | 2K on LaoZhang API | 2K on MiniMax API |
| --- | ---: | ---: | ---: | ---: |
| Text only, or first and last frame | $0.25 | $0.40 | $0.40 | $0.65 |
| 7 reference images | $0.31 | $0.48 | $0.46 | $0.73 |
| 9 reference images | $0.37 | $0.56 | $0.52 | $0.81 |
| 6-second reference video | $0.55 | $0.88 | $0.88 | $1.43 |
| 15 seconds of reference video | $1.00 | $1.60 | $1.60 | $2.60 |
| Reference audio | $0.25 | $0.40 | $0.40 | $0.65 |

The reference-video rows hide a rounding rule. LaoZhang API rounds each reference clip up to the next whole second, so a 5.8-second clip is billed as 6. MiniMax's pricing page doesn't say how it counts fractional reference seconds, so its column assumes whole-second clips. A real bill shows the rounding: on November 2, 2026, a 4-second 768P request with one 6.06-second reference video through LaoZhang API reported `input_seconds` 7 and `output_seconds` 4, and the balance dropped by exactly $0.55, or (4 + 7) × $0.05. Trimming that clip to 6.0 seconds would have saved $0.05. To keep a reference clip from being copied into the new scene, and for its limits of three clips and 15 seconds in total, see [How to Use a MiniMax H3 Reference Video Without Copying the Clip](https://blog.laozhang.ai/en/posts/minimax-h3-reference-video).

## Are failed or timed-out MiniMax H3 tasks charged?

On LaoZhang API you pay only for tasks that succeed. A request rejected at submission isn't charged, a task that fails is fully refunded, a task still `queued` after 30 minutes moves to `failed` and is refunded, and an HTTP 503 "The video service is busy" means no task was created and nothing was charged. MiniMax's pricing page doesn't say whether its own API bills failed tasks.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

![How a MiniMax H3 task is charged on LaoZhang API: requested seconds charged at creation, reference seconds billed on success, a full refund on failure, and no charge for HTTP 400 or 503 rejections](https://blog.laozhang.ai/posts/en/minimax-h3-api-pricing/img/task-billing-lifecycle.webp)

Both refund paths were exercised on November 2, 2026. Four malformed requests (four reference videos, a reference video mixed with a first frame, no text item, and reference audio as the only reference) came back as HTTP 400 at submission, the last one with the message "This request is not charged," and the balance didn't move. Two other requests broke a limit that's only checked after acceptance: one carried two 9.47-second reference clips, 18.95 seconds against the 15-second cap, and the other sent a 1.5-second clip as Base64 data instead of an HTTPS link. Both returned a `task_id`, the balance dropped by $0.40 for the two 4-second pre-charges, both failed about 3 seconds later with a message calling it a temporary service error, and the balance returned to its starting value. The message said "temporary," but the cause was in the input, so check reference lengths and links before you submit instead of retrying.

The one way to pay twice is to resubmit a task that hasn't failed. A client-side timeout isn't a failure: the task keeps running, and a new request with the same body is a new task billed again. Keep polling the original `task_id`, store it next to your own job ID, and recreate the request only after the task reports `failed`.

## How long a MiniMax H3 clip takes: 45 to 103 seconds in tests

Expect roughly one to two minutes per clip at the lengths tested so far. LaoZhang's docs say a 4–5 second 768P clip usually finishes in about a minute, that 2K and reference-video tasks take longer, and that clients should allow 30 minutes in total. In the November 2, 2026 tests, the 4-second 768P task with a reference video first reported `succeeded` on the poll at 45 seconds. It came back as a 4.46-second, 1344×768 clip at 24 fps with a stereo soundtrack.

Earlier the same day, five more `MiniMax-H3` clips through LaoZhang API took 50–103 seconds each and cost about $2.28 at list price (6 × $0.05 + 6 × $0.08 + 30 × $0.05). The times weren't recorded per clip, so they don't show how much longer 2K takes than 768P, and no 15-second clip was timed. The files are permanent, so you can judge the quality for the price yourself:

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

| Clip | Settings | Prompt |
| --- | --- | --- |
| [Fox at 768P](https://media.laozhang.ai/samples/video/h3-fox-768p-6s.mp4) | 768P, 6 seconds, 16:9 | A red fox walking through snowy forest at sunrise, cinematic |
| [Fox at 2K](https://media.laozhang.ai/samples/video/h3-fox-2k-6s.mp4) | 2K, 6 seconds, 16:9 | Same as the 768P fox |
| [Paper boat](https://media.laozhang.ai/samples/video/h3-paper-boat-768p-10s.mp4) | 768P, 10 seconds, 16:9 | A paper boat drifts down a rainy city gutter, then sails into a small sunny pond |
| [Hot air balloon](https://media.laozhang.ai/samples/video/h3-hot-air-balloon-768p-10s.mp4) | 768P, 10 seconds, 16:9 | A hot air balloon rises over a valley at dawn, drifts over a river and a village, then lands in a meadow |
| [Cyclist](https://media.laozhang.ai/samples/video/h3-cyclist-768p-10s-9x16.mp4) | 768P, 10 seconds, 9:16 | A cyclist rides through a coastal town from morning to sunset: market, harbor, cliff road, lighthouse |

## Does a 768P draft before a 2K final save money?

Yes, as soon as you throw away even one draft. Every 768P attempt costs $0.03 per second less than a 2K attempt on LaoZhang API, and $0.05 less on MiniMax's API. For a 10-second shot that takes five tries, four 768P drafts plus one 2K final cost $2.80 on LaoZhang API, against $4.00 for five 2K attempts.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

![Cost of one 10-second MiniMax H3 shot on LaoZhang API: five 2K attempts $4.00, four 768P drafts plus a 2K final $2.80, with a second 2K attempt $3.60, five 768P attempts $2.50](https://blog.laozhang.ai/posts/en/minimax-h3-api-pricing/img/draft-then-2k-cost.webp)

The catch is that the 2K render is a new generation, not an upscale of the draft you liked. The request has no seed field, so the 2K version can differ, and a second 2K attempt brings the same example to $3.60, still below $4.00. If you only need one good take and don't care whether it's 2K, finishing at 768P keeps the whole job at $2.50. MiniMax sells a separate Regeneration model that turns a 768P result into 2K for $0.05 per second, with the original inputs billed again, but it's only on MiniMax's own API.

## MiniMax H3 Max, Highspeed and Lite: official prices only

These are separate MiniMax models, not lighter settings of `MiniMax-H3`, and LaoZhang API doesn't serve them. As of November 12, 2026, MiniMax lists H3-Highspeed at $0.06 per second at 768P and $0.04 at 480P. H3-Lite is MiniMax's own low-cost lightweight model, and its per-second rates are on [MiniMax's pay-as-you-go pricing page](https://platform.minimax.io/docs/guides/pricing-paygo). H3-Max costs $0.08 per output second at 768P but $0.143 per second of input video, with only the first two images free and $0.074 per image after that. All three top out at 768P, so 2K output needs the full `MiniMax-H3`.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

If you are still deciding whether H3 is the model to test first, [Wan 3.0 vs MiniMax H3 vs Seedance 2.5: Pick the First Model to Test](https://blog.laozhang.ai/en/posts/wan-3-vs-minimax-h3-vs-seedance-2-5) compares the three. For plan and API costs across video generators, see [How Much Does an AI Video Generator Cost Per Month and Per Clip?](https://blog.laozhang.ai/en/posts/how-much-does-ai-video-generator-cost), and for [what a 10-second clip costs on each video API](https://blog.laozhang.ai/en/posts/cheapest-video-generation-api), see the per-clip comparison.

## Does running H3's open weights yourself cost less than the API?

Only if you already own the hardware and 768P is enough. The open H3 weights output 768p at 24 fps for 4–15 seconds; the 2K module isn't open-sourced, so 2K needs MiniMax's hosted APIs. The SGLang serving guide's example setups are 4×H200, 4×H100, or 2×RTX 5090 with 384 GiB of host RAM, plus a single RTX 4090 recipe at about 18 GB peak that was tested in first-frame mode, not reference mode. In that guide's own example, a reference-video job took 260 seconds against 37 seconds for text-to-video. Self-hosting makes sense when GPUs of that class already sit idle in your stack and 768P is your final resolution; otherwise, weigh your hardware bill against $0.25 per 5-second 768P clip.

## MiniMax H3 video API: make a first call and find the charge

A first test costs $0.25 and takes about a minute. These steps use cURL and `jq`:

1. In [LaoZhang API token management](https://api.laozhang.ai/token), create a token with group `default` and billing mode Usage first. Per-call tokens can't call this model, and you don't need a MiniMax account.
2. Send the create request below. It returns only a `task_id`.
3. Query the task every 10 seconds until `task.status` is `succeeded` or `failed`. If your client gives up after 30 minutes, keep the `task_id` and query it later instead of resubmitting.
4. Download the MP4 from `task.content.url`. The link needs no `Authorization` header and lasts 7 days. In the November 2, 2026 test, the download host returned 403 to Python urllib's default User-Agent while cURL worked, so set your own User-Agent in Python.
5. Open the [LaoZhang API call logs](https://api.laozhang.ai/log) and find the charge. A reference-video task shows a second entry when it finishes, because the reference seconds are settled then.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

```bash
export LAOZHANG_API_KEY="your-laozhang-api-token"

TASK_ID=$(curl -s -X POST "https://api.laozhang.ai/hailuo/v2/video_generation" \
  -H "Authorization: Bearer $LAOZHANG_API_KEY" \
  -H "Content-Type: application/json" \
  --data-raw '{
    "model": "MiniMax-H3",
    "content": [
      { "type": "text", "text": "A red fox walking through snowy forest at sunrise, cinematic" }
    ],
    "resolution": "768P",
    "duration": 5,
    "ratio": "16:9"
  }' | tee h3-create.json | jq -r '.task_id')
echo "task_id: $TASK_ID"

# Poll every 10 seconds for up to 30 minutes
for i in $(seq 1 180); do
  STATUS=$(curl -s "https://api.laozhang.ai/hailuo/v2/query/video_generation/$TASK_ID" \
    -H "Authorization: Bearer $LAOZHANG_API_KEY" | tee h3-task.json | jq -r '.task.status')
  echo "$(date +%T) $STATUS"
  case "$STATUS" in succeeded|failed) break ;; esac
  sleep 10
done

jq '{status: .task.status, usage: .task.usage, error: .task.error}' h3-task.json
[ "$STATUS" = "succeeded" ] && curl -L -o h3-output.mp4 "$(jq -r '.task.content.url' h3-task.json)"
```

If `task_id` prints as `null`, the request was rejected at submission and not charged; `h3-create.json` holds the error. For this 5-second text-only clip, `task.usage` should read `output_seconds` 5 and `input_seconds` 0, which matches a $0.25 entry in the call logs. Keep `resolution` uppercase (`768P` or `2K`) and `duration` a whole number, since lowercase or decimal values are rejected.

## Why LaoZhang API's MiniMax H3 price differs from MiniMax's

LaoZhang API runs H3 in its own independent deployment, not through MiniMax's official API, so it has its own pricing. That separate deployment also explains the practical differences you'll meet when moving code between the two.

*Disclosure: [laozhang.ai is our own service](https://blog.laozhang.ai/en/about#laozhang). Its prices are platform pricing, not an independent comparison.*

The paths start with `/hailuo/v2/` on `https://api.laozhang.ai`, so code written for MiniMax's own API needs its base URL and paths changed. Video and audio inputs must be public HTTPS links, whereas MiniMax's API also accepts its own `mm_file://` IDs and Base64 video. MiniMax H3 isn't available on the OpenAI-style `/v1/videos` or `/v1/video/generations` endpoints, which reject such requests without charge.

## Sources

External pages this guide links to, in the order they appear. Last updated 2026-11-12.

- [MiniMax's pay-as-you-go pricing page](https://platform.minimax.io/docs/guides/pricing-paygo) (platform.minimax.io)