MiniMax H3-Image Price: Cost per Image at Every Provider
I paid 4.98 USD for 170+ images on launch day. Here is MiniMax H3-Image pricing at every provider, and the pixel trap that doubled four of my 4K bills.

👁 111 reads
MiniMax H3-Image costs between 2 and 9 cents per image through the API, depending on where you buy it and at what resolution. On launch day I paid 4.98 USD for more than 170 images. What taught me the most, though, was four portraits that cost me twice as much as the 4K images next to them, even though they looked the same. Below are the prices at every provider I checked on 10 November 2026, and how to avoid overpaying.
MiniMax H3-Image price: you pay for tokens
The official API (platform.minimax.io) has no price “per image”. It counts tokens, like a language model:
| Model | Text input | Image input | Image output |
|---|---|---|---|
| MiniMax-H3-Image | 2.30 USD / M tokens | 1.50 USD / M | 3.80 USD / M |
| MiniMax-H3-Image-Highspeed | 2.30 USD / M | 3.00 USD / M | 7.60 USD / M |
| H3-Image-Context-IR (prompt expander only) | 2.30 USD / M | - | - |
MiniMax gives examples converted to a per-image price: 1K 0.02179 USD, 2K 0.02333 USD, 4K 0.06879 USD. An edit with one reference at 4K is 0.07312 USD, with three references 0.07914 USD. Highspeed: 1K 0.03529, 2K 0.03622 USD. On top of that there is a promotion: 50% off until 17 November 2026, 16:00 UTC, but only for images of at least 16,600,000 pixels. With it, a 4K image costs 0.03439 USD.
The 16.6 million pixel trap
In my test I made several 4K images. Landscapes at 5472×3072 and squares at 4096×4096 cost 0.033-0.036 USD. Four portraits in 4:5 proportions, which I set to 3648×4544, cost 0.067-0.070 USD. I counted the pixels: 3648 x 4544 = 16,576,512. That is 23,488 pixels short of the promo threshold, so I paid the full rate.
| Size | Pixels | Promo? | My cost |
|---|---|---|---|
| 5472×3072 (16:9) | 16,809,984 | yes | 0.033-0.035 USD |
| 4096×4096 (1:1) | 16,777,216 | yes | 0.034-0.036 USD |
| 3648×4544 (4:5) | 16,576,512 | no | 0.067-0.070 USD |
| 3680×4512 (4:5) | 16,604,160 | yes | 0.033-0.035 USD |
The lesson: before you order a 4K image, multiply width by height. The easiest route is to take the sizes from the MiniMax table (4096×4096, 4736×3552, 5472×3072, 6272×2688) or use the area: "4K" parameter instead of your own dimensions.
Price comparison: every provider (as of 10 November 2026)
| Provider | 2K | 4K | Editing | Notes |
|---|---|---|---|---|
| MiniMax API | 0.02333 USD | 0.03439 USD (promo) | from approx. 0.027 USD | tokens, up to 10 images per request, synchronous or streamed |
| fal.ai | tokens, same as MiniMax | yes | Standard and Highspeed | |
| WaveSpeedAI | 0.03 USD | 0.09 USD | from 0.038 USD | flat price, no promo |
| Hailuo AI | 3 credits per image | yes | app, not API | |
| OpenRouter, Replicate | not available | - | - | |
I also ran one 4K image through WaveSpeedAI: 5472×3072, 23.6 s, of which 19.7 s was pure generation. Going by its price list I paid 0.09 USD, almost three times what MiniMax charges with the promo. After 17 November the gap will shrink to about 30%.
What my tests cost
| Type | Time | Cost from the cost field |
|---|---|---|
| 2K with prompt expander (default) | 13-17 s | 0.020-0.027 USD |
| 2K without expander | 12-16 s | 0.0156 USD |
| 4K (from 16.6 MP) | 22-24 s | 0.033-0.036 USD |
| Highspeed 2K | 12.7-15.3 s | 0.035-0.039 USD |
| Edit, 1 reference, 2K | 20-27 s | 0.032-0.034 USD |
| Edit, 3 face photos, 2K | 25-32 s | 0.036-0.044 USD |
| Edit, 16 references | 56-68 s | 0.098-0.099 USD |
Two things surprised me. First, the prompt expander costs about 0.006 USD per image, because its tokens go onto the bill. If you already have a polished, long prompt in English, you can switch it off (disable_prompt_enhanced: true) and pay 0.0156 instead of roughly 0.023 USD. In my test of Polish text with diacritics, the result without the expander was even slightly better (21 errors instead of 24-29). Second, Highspeed at 2K was no faster than the standard version, yet cost 1.6 times more.
How to use the MiniMax H3-Image API
The shortest text-to-image request in Python looks like this (the key from your MiniMax account goes in an environment variable):
import os, requests
r = requests.post(
"https://api.minimax.io/v1/images/generations",
headers={"Authorization": "Bearer " + os.environ["MINIMAX_API_KEY"]},
json={"model": "MiniMax-H3-Image",
"prompt": "Luxury product shot of a steel diver watch on wet black slate",
"size": "5472x3072", "n": 1, "response_format": "url"},
timeout=600)
for i, im in enumerate(r.json()["data"]):
print(i, im.get("url") or im.get("error"))
Editing goes through /v1/images/edits with an images list (1-16 images, each up to 64 MB, up to 64 MiB in total per request). The order in the list gives you “Picture 1”, “Picture 2” and so on, and those are the names you use to refer to them in the prompt. The response has a usage field with token counts, so you can work out the cost yourself. One warning from the documentation: the request is not idempotent, so in the OpenAI SDK client turn off automatic retries, or you may pay twice.


Which option to choose
- Until 17 November: the official MiniMax API or fal.ai, with 4K images at the sizes from the table. I found nothing cheaper at 4K.
- After the promo: 2K at 0.023 USD is cheaper than Seedream 5.5 at 2K on WaveSpeedAI (0.035 USD). 4K goes back to 0.069 USD, so it is worth checking whether 2K is enough.
- No coding: Hailuo AI, 3 credits per image.
Quality results and comparisons with rivals are in the MiniMax H3-Image hub.
Sources
- MiniMax: Pricing Overview (10 November 2026)
- MiniMax: Image Generation Guide
- MiniMax: Text-to-Image API
- WaveSpeedAI: MiniMax H3 Image pricing (10 November 2026)
- fal.ai: MiniMax H3 Image
- My own test on 10 November 2026: more than 170 images through the API, costs taken from the cost field of the response; 1 image through WaveSpeedAI.
Checked: 10 November 2026.
More on this from QUANTSLANT
guides·24 Tested MiniMax H3-Image Prompts With Results24 prompts from very different worlds, each run once in MiniMax H3-Image, with the full prompt, time and cost under every image.