
The MAI Image 2.7 API replaces MAI Image 2.6 as Microsoft's newest image model on PoYo, with larger maximum sizes, transparent backgrounds and a choice of PNG, JPEG or WebP. The one thing it drops is web search grounding. This post compares mai-image-2.7 with mai-image-2.6 field by field and shows the exact JSON changes to migrate.
Short answer: move to MAI Image 2.7 for new work and most existing traffic, especially Flash traffic, whose output rate drops by about a third. Keep MAI Image 2.6 only for requests that depend on enable_web_search.
Release and positioning
| MAI Image 2.6 | MAI Image 2.7 | |
|---|---|---|
| Microsoft release | Announced August 10, 2026; Flash added September 4, 2026 | November 17, 2026, with Flash |
| PoYo model IDs | mai-image-2.6, mai-image-2.6-flash | mai-image-2.7, mai-image-2.7-flash |
| Microsoft Foundry status | Listed without a preview label | Public preview |
| Microsoft's description | "Stronger text rendering, improved portraits and 3D imagery" | "Precise image editing, multi-reference editing, improved text rendering, high-fidelity portraits" |
Microsoft says MAI Image 2.7 was "rebuilt from the ground up for multimodal editing workflows" and that it ranked #1 on Artificial Analysis's AA-Image-Editing v2.1 leaderboard as of November 16, 2026 (1,215 Elo across 5,478 votes). Microsoft's Foundry documentation lists transparent background as a capability 2.7 adds over the rest of the family.
On Foundry, "preview" means no service-level agreement. On PoYo both generations are callable with the same key and the same async task flow.
MAI Image 2.7 API vs MAI Image 2.6 API: price per token
PoYo rates for each model. 1 credit = $0.005.
| Per 1M tokens | MAI Image 2.6 | MAI Image 2.7 |
|---|---|---|
| Text input | 800 credits ($4.00) | 1,000 credits ($5.00) |
| Image input | 1,280 credits ($6.40) | 1,600 credits ($8.00) |
| Image output | 6,080 credits ($30.40) | 6,800 credits ($34.00) |
| Flash text input | 280 credits ($1.40) | 350 credits ($1.75) |
| Flash image input | 400 credits ($2.00) | 500 credits ($2.50) |
| Flash image output | 3,040 credits ($15.20) | 2,000 credits ($10.00) |
Price per same request
On PoYo, output tokens are about one per 1,024 pixels on both generations, so cost tracks pixel count. Output-token cost only:
| Request | MAI Image 2.6 | MAI Image 2.7 | 2.6 Flash | 2.7 Flash |
|---|---|---|---|---|
1:1, 1K (1024x1024 on both) | 6.23 credits ($0.0311) | 6.96 credits ($0.0348) | 3.11 credits ($0.0156) | 2.05 credits ($0.0102) |
1:1, 2K | 1536x1536: 14.01 credits ($0.0700) | 2048x2048: 27.85 credits ($0.1393) | 7.00 credits ($0.0350) | 8.19 credits ($0.0410) |
16:9, 2K | 2048x1152: 14.01 credits ($0.0700) | 2720x1536: 27.74 credits ($0.1387) | 7.00 credits ($0.0350) | 8.16 credits ($0.0408) |
| Per megapixel | $0.0297 | $0.0332 | $0.0148 | $0.0098 |
Two things to notice:
- 2.7 Flash has the lowest output rate of the four. At the same pixel count it costs about a third less than 2.6 Flash. The flagship 2.7 is the tier for precise edits, multi-reference work, transparency and sizes up to 4 megapixels; 2.6 is the tier for web-grounded prompts.
2Know means more pixels. MAI Image 2.6's 2K is about 2.36 megapixels; MAI Image 2.7's is about 4. If your code sends"resolution": "2K"to the flagship, each image uses about 1.8 times the output tokens of 2.6's 2K (4,080 vs 2,304 at 16:9, about 27.7 vs 14.0 credits), because you get roughly 1.8 times the pixels. To keep 2.6's 2K pixel count on 2.7, send a custom size such as2048x1152or1536x1536: 2,304 output tokens, about 15.67 credits ($0.078) on MAI Image 2.7.
On launch day (PoYo production API, November 19, 2026) a 2K 16:9 request returned 2720x1536 on both 2.7 tiers; at current rates it comes to 27.802 credits on mai-image-2.7 and 8.1803 credits on mai-image-2.7-flash. Full cost tables are in the MAI Image 2.7 API pricing post.
Field-by-field request differences
| Field | MAI Image 2.6 API | MAI Image 2.7 API |
|---|---|---|
model | mai-image-2.6 / mai-image-2.6-flash | mai-image-2.7 / mai-image-2.7-flash |
prompt | 1 to 32,000 characters | Same |
image_urls | 1 to 5 URLs; output about 1 MP and its size follows the reference images | 1 to 5 URLs; without size, follows the first reference's aspect ratio at about 1 MP |
size with image_urls | Ignored (must still be valid) | Applied: sets the output ratio or exact size |
size presets | 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16 | Same |
Custom WIDTHxHEIGHT | Sides divisible by 32, each at least 768, at most 2,359,296 pixels | Sides divisible by 16, each at least 768, at most 4,194,304 pixels |
resolution: "2K" | Fixed table, e.g. 1536x1536, 2048x1152 | About 4 MP, e.g. 2720x1536 at 16:9 |
enable_web_search | Supported (default false) | Not accepted |
background | Not available (opaque only) | opaque (default) or transparent |
output_format | Not available (PNG only) | png (default), jpeg, webp |
Both schemas reject unknown fields, so a leftover enable_web_search makes a MAI Image 2.7 request fail validation.
Migration diff
Most migrations are a model ID swap plus removing web search:
{
- "model": "mai-image-2.6",
+ "model": "mai-image-2.7",
"input": {
"prompt": "A travel poster of the Eiffel Tower at night with the text \"Paris\"",
"size": "2:3",
- "enable_web_search": true
+ "resolution": "2K",
+ "output_format": "webp"
}
}
The same change applies to Flash: mai-image-2.6-flash becomes mai-image-2.7-flash.
Checklist before you switch production traffic:
- Remove
enable_web_search. If a prompt relied on it for current facts, write those facts into the prompt yourself, or keep that request on MAI Image 2.6. - Review
sizeon edit requests. MAI Image 2.6 ignoredsizewhenimage_urlswas set; MAI Image 2.7 applies it. If your client always sends a default like"1:1", edits of landscape photos will now come back square. Omitsizeon edits to keep the source ratio. - Check
2Kbudgets. 2K on the flagship returns about 1.8 times the pixels and costs more per image; use a custom size if you need the old dimensions. - Custom sizes carry over. Every valid MAI Image 2.6 custom size (multiples of 32, at most 2,359,296 pixels) is also valid on MAI Image 2.7.
- Opt into the new fields where useful.
"background": "transparent"with PNG or WebP for cutouts, and"output_format": "jpeg"or"webp"for smaller photographic files. See the transparent background guide.
Polling, callbacks, status fields and error codes are the same for both generations, so no other client code changes. The MAI Image 2.7 API guide has a full Python client.
When to stay on MAI Image 2.6
- Web-grounded prompts: current events, recently opened places, new products. MAI Image 2.7 has no equivalent field.
- A pinned visual style. A new model generation produces different images from the same prompt. If you have approved outputs from 2.6, compare a sample on 2.7 before switching.
For a broader comparison of MAI Image 2.6 against other image APIs, see MAI Image 2.6 vs GPT Image 2.6 vs Nano Banana 2.1.
Frequently asked questions
How do MAI Image 2.7 API prices compare with the MAI Image 2.6 API? On PoYo, MAI Image 2.7 is $5.00 / $8.00 / $34.00 per 1M text input, image input and output tokens, and MAI Image 2.6 is $4.00 / $6.40 / $30.40. On Flash, 2.7 has the lower output rate, $10.00 vs $15.20 per 1M, so a 1024x1024 Flash image costs about $0.0102 instead of $0.0156. Choose 2.7 for editing, transparency and larger sizes, and 2.6 for web search grounding.
Does the MAI Image 2.7 API support web search like MAI Image 2.6?
No. enable_web_search is not part of the MAI Image 2.7 schema and is rejected.
Do I need to change my code to move from the MAI Image 2.6 API to MAI Image 2.7?
Usually only the model ID and removing enable_web_search. Also check edit requests that send size, because 2.7 applies it.
What is the largest image the MAI Image 2.7 API can produce compared with 2.6? 4,194,304 pixels on 2.7 versus 2,359,296 on 2.6, about 1.8 times more.
Can MAI Image 2.6 make transparent images? No. Transparent backgrounds and JPEG or WebP output are only on the MAI Image 2.7 API.