MiniMax H3 Image API: Generation, Editing and Real Pricing

Same-model provider prices: Cheap MiniMax H3 Image API · 2026-11-13

MiniMax H3 Image capabilities and hosted API limits

The strongest reason to evaluate H3 Image is its unified generation-and-editing workflow: describe a new asset, supply product or character references, then ask for a targeted change. MiniMax also publishes weights for local deployment. Published capability claims are not a guarantee of exact logos, typography or identity preservation in every output.

Hosted API feature	Documented contract
Generation	POST /v1/images/generations
Editing	POST /v1/images/edits, 1–16 reference images
Prompt	Up to 32,000 characters; without enhancement, only about the first 4,096 tokens take effect
Output dimensions	655,360–20,000,000 pixels; aspect ratio from 1:3 to 3:1
Output count	n from 1 to 10
Formats	JPEG by default; PNG or WebP supported
Responses	URL or non-streaming Base64; native SSE supported
Unsupported controls	Masked editing, transparent backgrounds and preview images

Reference files must meet MiniMax's input limits: JPG/JPEG, PNG, WebP, HEIC or HEIF, each dimension 128–8,192 pixels, long-to-short ratio at most 4:1, and total request body at most 64 MiB. Base64 increases request size. Use accessible image URLs or suitably sized uploads; the model's ability to accept 16 images does not remove the body-size limit.

“4K” is an area setting in this API. area: "4K" targets about 4096 × 4096 pixels, with dimensions derived from aspect ratio. At 16:9, the documented output is 5472 × 3072, not 4096 × 2304. For explicit dimensions, send size: "WIDTHxHEIGHT" and omit area and aspect_ratio. For area-based sizing, use size: "auto", area and aspect_ratio. Check data[].size and the downloaded image's actual dimensions.

How much does MiniMax H3 Image cost?

For the current APIMaster Standard route, the regular-price calculation is:

USD = (text_input_tokens × 2.30
     + image_input_tokens × 1.50
     + image_output_tokens × 3.80) / 1,000,000

Use separate text and image usage fields; do not charge all input tokens as text and then add image input again. Prompt enhancement contributes billed text tokens, so resolution alone does not determine the total price. Account-specific coefficients, where applicable, and billing rounding can affect the final debit.

MiniMax's official illustrative Standard totals show the scale of the bill. These are examples, not APIMaster fixed tariffs or our measured average:

Task	1K: 1024 × 1024	2K: 2048 × 2048	4K: 4096 × 4096, regular rate
Text-to-image	$0.02179	$0.02333	$0.06879
Edit with one 1K reference	$0.02654	$0.02658	$0.07312
Edit with three 1K references	$0.03563	$0.04177	$0.07914

For context, 1,000 text-to-image requests at the first example would be approximately $21.79; that is an illustrative token-budget calculation, not a prepaid package. Actual prompts and enhancement can change it. Nano Banana 2.1 has a different, resolution-based APIMaster contract, starting at $15.12 for 1,000 1K images before any separately quoted conditions.

Does the official 50% 4K promotion apply on APIMaster?

Do not assume that it does. MiniMax advertises 50% off the total request cost for Standard and Highspeed until November 17, 2026, 16:00 UTC, exclusive—November 18 at 00:00 Beijing time—when the final delivered image has at least 16,600,000 pixels. Its square 4K text-to-image example changes from $0.06879 to $0.03439 during that offer.

A 4096 × 4096 output qualifies by area; 4096 × 2304 does not. A “4K” label alone is insufficient. The current APIMaster public route uses regular token rates, without this temporary promotion. Compare the actual checkout/usage contract, not MiniMax's campaign headline. After the deadline, the promotional examples should not be used as current prices.

MiniMax H3 Image API example: text-to-image

Create an APIMaster account, add balance for paid image requests, then create a key in API keys. Set APIMASTER_API_KEY in your environment. This example uses the Standard route's exact, case-sensitive model ID:

import os
import requests

base_url = "https://apimaster.ai/v1"
headers = {"Authorization": f"Bearer {os.environ['APIMASTER_API_KEY']}"}
response = requests.post(
    f"{base_url}/images/generations",
    headers=headers,
    json={
        "model": "MiniMax-H3-Image",
        "prompt": "A studio product photo of a matte blue teapot and two pears on a cream table, soft daylight, no text.",
        "size": "auto",
        "aspect_ratio": "16:9",
        "area": "2K",
        "n": 1,
        "response_format": "url",
    },
    timeout=600,
)
response.raise_for_status()
result = response.json()
for item in result.get("data", []):
    if item.get("error"):
        print("Image failed:", item["error"])
    else:
        print(item.get("url"), item.get("size"))
print("Usage:", result.get("usage"))

This is a request example, not an image-quality benchmark. Download successful output URLs promptly and store the files yourself. Check per-image errors: with n > 1, HTTP 200 can include failed images. A timeout does not prove a request was rejected. H3 generation is not idempotent, so an automatic retry can create another paid generation. If using the OpenAI SDK, set a sufficiently long timeout and max_retries=0.

MiniMax H3 Image editing: reference images and prompt control

For JSON editing, send reference URLs in images[].image_url. Do not copy Nano Banana's image_urls field into this H3 request. Replace the example URLs with your own accessible assets:

edit = requests.post(
    f"{base_url}/images/edits",
    headers=headers,
    json={
        "model": "MiniMax-H3-Image",
        "prompt": "Keep the product shape and label from image 1. Use the background palette from image 2. Do not add text or change the product.",
        "images": [
            {"image_url": "https://example.com/product.png"},
            {"image_url": "https://example.com/palette.png"},
        ],
        "size": "2048x2048",
        "n": 1,
        "response_format": "url",
    },
    timeout=600,
)
edit.raise_for_status()
print(edit.json())

APIMaster's existing November 11 integration acceptance checks covered HTTP references, data URIs, multipart uploads, multi-reference editing, SSE and 1K/2K/4K decoded outputs. A controlled two-reference test retained four colored shapes and two short labels. These were interface and billing checks with simple assets, not a broad photographic or typography evaluation; they do not establish superiority over another model.

Generation uses prompt enhancement by default; you can disable it with disable_prompt_enhanced: true while supplying fixed dimensions or a fixed aspect ratio. Edits always use enhancement. Inspect revised_prompt when diagnosing unexpected changes. For product work, specify what must stay fixed, identify each reference's role and verify the rendered label before publishing it.

Standard vs Highspeed, and open weights vs hosted API

Highspeed uses the same documented API parameters, but its official image-input and image-output rates are twice Standard's; text input stays $2.30/M. APIMaster currently lists Standard only. MiniMax's launch speed claims depend on model variant, size and deployment and are not a measured APIMaster latency guarantee.

The official Hugging Face repository provides weights and integrations including Diffusers, SGLang, vLLM-Omni and ComfyUI. Its BF16 checkpoint is about 68.65 GB, with a separate approximately 66.7 GB text encoder. Download size is not a complete GPU-memory estimate. Local weights use the MiniMax H3 Image community license, so review that license before deployment; “open weights” does not mean unrestricted use under MIT or Apache.

Use hosted access to evaluate prompts without building inference infrastructure. Consider local deployment when control over inference justifies GPU, maintenance and licensing work. Local pipeline limits and hosted API limits are different contracts.

FAQ

Is MiniMax H3 Image the same as MiniMax H3 video?

No. H3 Image generates and edits still images using MiniMax-H3-Image. The H3 video family has different model IDs, endpoints and billing. See the separate MiniMax H3 video guide.

Is MiniMax H3 Image available on APIMaster?

Yes. The November 13 check found one enabled public Standard route, billed at $2.30/M text input, $1.50/M image input and $3.80/M image output. Highspeed had no public route. Check the live model card before sending requests.

Is H3 Image a fixed $0.02-per-image API?

No. Official per-image examples are illustrative totals. Actual token usage, prompt enhancement, reference inputs and output dimensions determine the bill. The current APIMaster route uses the regular token formula.

Can H3 Image edit with 16 references or a mask?

It documents up to 16 reference images within format and request-size limits. Masked editing and transparent backgrounds are not supported. Reference-based instructions do not guarantee pixel-exact preservation.

Which is cheaper: H3 Image or Nano Banana 2.1?

For the illustrated 1K text-to-image task, H3's regular official example is $0.02179, versus APIMaster Nano Banana 2.1 from $0.01512 per 1K image. This compares an example token total with a resolution quote; measure actual accepted-output cost for your prompts. See the dedicated comparison.

Sources and next steps

Sources checked November 13, 2026: MiniMax release, hosted API guide, official pricing and promotion, model weights and license, Google image pricing, and APIMaster live market cards. Capability descriptions come from their respective vendors; price checks and the dated APIMaster integration observations are identified separately.

Open the H3 Image card, register, create an API key, and send MiniMax-H3-Image to https://apimaster.ai/v1. For a discounted alternative with a different billing contract, compare Nano Banana 2.1 routes, currently up to 55% below Google's image-output reference.