MAI Image API: How to Call Microsoft's MAI-Image-2.6 and 2.6 Flash, Every Route and What an Image Costs (2026)

Microsoft sells MAI-Image-2.6 and MAI-Image-2.6 Flash through Microsoft Foundry, in public preview since September 4, 2026, priced by token: $38 per million image output tokens on 2.6 and $19 on Flash, plus a few dollars per million for the prompt and any picture you send. That comes to about 4 cents for a 2.6 image at launch, by RuntimeWire's estimate, and about half that on Flash. Without an Azure account, OpenRouter and Vercel's gateway pass Microsoft's token prices through, one reseller sells a flat price per image, and GenMagic runs both models beside every other image and video model on one balance.

This guide covers the MAI Image models, what Microsoft charges and what that means per image, getting access through Foundry, every route without an Azure account, working requests, what MAI Image did in our tests, limits, and which route to pick.

Not a developer? To make MAI Image pictures in a studio and pay per image, without code or a subscription, see MAI Image pay as you go.

A note on who wrote this: GenMagic is one of the routes compared here. Its prices come from its live catalog. Microsoft's facts come from its Foundry announcement and the MAI-Image-2.6 model card; other platforms' prices were read on their own pages on November 10, 2026.

Table of contents

1. The short answer
2. The MAI Image models
3. What Microsoft charges
4. Getting access through Microsoft Foundry
5. MAI Image without an Azure account
6. Calling MAI Image through GenMagic
7. What MAI Image did in our tests
8. Limits and gotchas
9. Which route to pick
10. How we checked
11. Questions

The short answer

• Two current models: MAI-Image-2.6, Microsoft's quality model, and MAI-Image-2.6 Flash, "an optimized version" for speed and volume. Both make and edit images, with up to five reference pictures in one request.
• Microsoft's price is per token: $38 per million image output tokens on 2.6 and $19 on Flash; text input costs $5 and $1.75 per million, image input $8 and $2.50.
• Per image that is roughly 4 cents on 2.6 and 2 cents on Flash for a text-only request; a picture you send to edit adds input tokens.
• Access from Microsoft is a Foundry project on an Azure subscription; the models are in public preview, so a new subscription can start without quota for them.
• Without an Azure account, OpenRouter and Vercel's AI Gateway charge Microsoft's token prices, and APIYI sells a flat $0.12 and $0.06 per image.
• On GenMagic the two cost $0.058 and $0.037 per image today, with one balance for every model.
• In our tests MAI-Image-2.6 set every word right on a text-heavy ad and made one of the most legible YouTube thumbnails; it also added glasses to a woman who does not wear them.

The MAI Image models

Release dates are the model cards'. Microsoft's card for 2.6 describes a diffusion model "designed for both high-quality text-to-image synthesis and precise, controllable image-to-image editing", with "improved rendering of text within generated images, including labels, posters, packaging, and signage". At launch, Microsoft says, MAI-Image-2.6 ranked second on Arena's text-to-image leaderboard and third on its image-editing leaderboard.

Image size. The largest image has 2,359,296 pixels, the area of 1536 by 1536, and "either dimension can exceed 1536 as long as the total stays within the limit".

What Microsoft charges

Foundry bills MAI Image by token, at these list prices from Microsoft's announcement (per million tokens):

Microsoft does not publish how many tokens one image counts. RuntimeWire put MAI-Image-2.6 at "about 4 cents an image" on launch day, which matches roughly a thousand output tokens per image at $38 per million, and Flash's output price is half. An edit costs more than a picture from text alone, because the photo you send is billed as image input: in our tests, a text-only MAI-Image-2.6 image was billed 5.1 cents on GenMagic and an edit of a room photo 9.5 cents ( image test, virtual staging test).

Getting access through Microsoft Foundry

1. An Azure subscription. A pay-as-you-go subscription works; a free or trial subscription may have no quota for new image models.
2. A Foundry project at ai.azure.com, where MAI-Image-2.6 and 2.6 Flash are in the model catalog, marked as preview.
3. Deploy the model in the project; its page then shows the endpoint and the key your code sends.
4. Quota. Because the models are in preview, a new subscription may need its quota approved before the first call works.

The MAI-Image-2.6 model card names two channels: Foundry for developers, and MAI Playground, a public site where anyone can try the models with low daily limits. Foundry's terms are Azure's, and output goes through Microsoft's prompt and output filters, which "help detect and block harmful, abusive, or policy-violating content".

MAI Image without an Azure account

Read on each platform's own page on November 10, 2026; GenMagic's prices are live:

Sources: OpenRouter, 2.6 and Flash, Vercel AI Gateway, 2.6 and Flash, APIYI. OpenRouter adds a fee when you buy credits.

What separates them:

• OpenRouter and Vercel's gateway pass Microsoft's token prices through, with Azure as the provider. On OpenRouter the image endpoint takes eight aspect ratios (1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3 and auto), one image per request, and up to five reference pictures, and returns a base64 PNG.
• APIYI charges a flat price per image, any size, which is easier to budget than tokens and several times the token price for a square image. It sets the size in pixels (at least 768 a side) and advises a client timeout of at least two minutes for Flash and three for 2.6.
• GenMagic charges per image, with the price shown before each generation; MAI Image runs beside every other image, video and voice model on one balance, in a studio, n8n, Zapier, ComfyUI and Claude or ChatGPT.

Calling MAI Image through GenMagic

GenMagic's API is OpenAI-compatible, so MAI Image runs through the same images.generate call as every other image model, with no Azure account. The model ids are microsoft/mai-image-2.6, microsoft/mai-image-2.6-flash, microsoft/mai-image-2.5 and microsoft/mai-image-2.5-pro; every image model and its price is in the image API pricing list.

1. Get a key. Create a free account at genmagic.co/get-api-key and verify your email; the account starts with $0.50 of credit, about 13 MAI-Image-2.6 Flash images.

2. Generate. With OpenAI's Python SDK:

import os
from openai import OpenAI

client = OpenAI(base_url="https://genmagic.co/api/v1", api_key=os.environ ["GENMAGIC_API_KEY"])

result = client.images.generate(
    model="microsoft/mai-image-2.6",
    prompt='A vintage enamel shop sign on a brick wall reading exactly "MORNING CROWN BAKERY", straight-on photo, soft daylight',
    size="1024x1024",
)
print(result.data [0].url)

Or with curl, editing a picture you send:

curl https://genmagic.co/api/v1/images/generations \
  -H "Authorization: Bearer $GENMAGIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "microsoft/mai-image-2.6-flash",
    "prompt": "Keep the bottle and its label exactly; place it on a sunlit bathroom shelf with a eucalyptus sprig",
    "image": "https://your-cdn.com/product.jpg"
  }'

The answer carries a hosted url (or the bytes with "response_format": "b64_json"), and the response headers say what the image cost and what is left on the balance. A refused or failed image costs nothing.

What MAI Image did in our tests

We ran MAI-Image-2.6 beside eight other image models on the same briefs in October 2026:

• Exact text, including a long ad. It spelled a three-line shop sign correctly in our image generator test, and set every line of a text-heavy story ad exactly, check marks included, in our ad test, where seven of nine models managed that. Its "20% OFF" badge came out so small it barely read on a phone.
• Thumbnails that read at small size. In our thumbnail test it made one of the three most legible layouts, with a large title, a large face and a clear subject.
• Faces kept, details added. It kept both people recognizable in our headshot test, but read "keep their glasses" as an instruction and gave glasses to a woman who does not wear them.
• Smaller images. Its tall story ad was 768 by 1365 pixels, the smallest in the ad test, and its staged rooms 1365 wide in our virtual staging test, where most models returned 2048 or more. Upscale before print or a full-screen listing.
• Speed. About 21 to 25 seconds an image in every test, against 7 to 9 seconds for the fastest model and over a minute for the slowest.

Every image model in those tests runs on one GenMagic balance, so the same prompt can go to MAI Image and the others to compare: the pay-as-you-go AI image generator.

Limits and gotchas

One image per request. OpenRouter's image endpoint accepts n of 1 only; send several requests for several images.

Describe only what is there. In our headshot test the model added glasses after a prompt asked it to keep them. Name what a person actually has, and say what must not change.

Preview terms. Both 2.6 models are in public preview on Foundry, where quota, regions and prices can change before general availability.

Safety filters. Microsoft filters prompts and outputs for harmful content; its model card names violent, gory and sexual content among the risks. A blocked request on GenMagic is not charged.

Size limit. 2,359,296 pixels at most, the area of 1536 by 1536; a wide banner can be longer on one side as long as the total stays under it.

Which route to pick

• You build on Azure already: Foundry, at Microsoft's token prices on your Azure bill, once your subscription has quota for the preview models.
• You want MAI Image beside other labs' models on one key, at Microsoft's token prices: OpenRouter or Vercel's AI Gateway.
• You want a fixed price per image: APIYI, or GenMagic, where each image's price shows before you generate and the same balance pays for video and voice models too.
• You want the cheapest MAI image: MAI-Image-2.6 Flash, at half of 2.6's output price everywhere it is sold.

How we checked

Microsoft's prices, dates and channels were read in its Foundry announcement of September 4, 2026 and the MAI-Image-2.6 model card, and the other platforms' prices on their own pages on November 10, 2026, linked in the table. GenMagic's prices come from its live catalog, and the quality results from our published tests of October 29 and 30, 2026.

Questions

Is there a MAI Image API? Yes. Microsoft sells MAI-Image-2.6 and MAI-Image-2.6 Flash through Microsoft Foundry, in public preview since September 4, 2026, and OpenRouter, Vercel's AI Gateway, APIYI and GenMagic carry them without an Azure account.

How much does MAI-Image-2.6 cost? On Foundry, $38 per million image output tokens, $5 per million text input tokens and $8 per million image input tokens, about 4 cents for a picture from text. On GenMagic it costs $0.058 per image today.

What is the difference between MAI-Image-2.6 and 2.6 Flash? Flash is Microsoft's optimized version of 2.6, at half its output price, and faster: about 17 seconds an image against 30 for 2.6 at 1024 by 1024, by APIYI's measurements. 2.6 is the quality model.

Can I use MAI Image for free? MAI Playground lets anyone try the models with low daily limits. On GenMagic, a new account gets $0.50 of credit once its email is verified, about 13 Flash images.

Can MAI Image edit a photo? Yes. MAI-Image-2.6 takes up to five reference pictures in one request and is built for "precise, surgical edits" such as removing or replacing objects and updating text, by Microsoft's model card.