Key Takeaways
- Top of the image editing board: 1,215 Elo on Artificial Analysis AA-Image-Editing v2.1 — 4 points ahead of GPT Image 2.6 Sunburst and 64 points above the previous MAI-Image-2.6
- New provider specs: maximum output rises from 1536×1536 to 2048×2048 (4,194,304 pixels), transparent backgrounds are now supported, and edits accept up to five reference images
- Per-request pricing unchanged: $0.12 / $0.06 per image, the same for 2048 and 1024, and the same for one reference image or five
- Effective cost is lower: with the current top 20% top-up bonus, the flagship lands at about $0.10 per image and Flash at about $0.05
- Migration is a model-name change: endpoints, parameters and error behavior match 2.6; APIYI has verified 2048 output, transparent backgrounds and multi-reference edits in production
Background
MAI-Image 2.7 and MAI-Image-2.7-Flash are the newest image models from Microsoft AI, announced on November 17, 2026 at Microsoft Ignite and available in public preview on Microsoft Foundry. The line moves fast: MAI-Image-2.6 arrived in September 2026; an editing-only 2.7 Preview followed on October 27 (No. 3 on the Arena image editing board); the November 17 release adds text-to-image generation, a cheaper Flash variant, and Canvas, an interactive image workspace in MAI Labs powered by Flash. Microsoft says MAI Image is now the default image model behind Bing Image Creator. For APIYI users there are two concrete changes: maximum output rises from 1536×1536 to 2048×2048, and transparent backgrounds are now supported. Pricing stays at the same per-request rates as the previous generation.Release status: both models are in public preview; Microsoft has not announced a GA date. APIYI serves them through the official Microsoft channel, with model IDs
MAI-Image-2.7 and MAI-Image-2.7-Flash. Model names are case-sensitive.Deep Dive
Benchmarks
Artificial Analysis AA-Image-Editing v2.1, as of November 20, 2026:
Text-to-image is a different story. On AA-Image-T2I v2.1, 2.7 ranks No. 4 (1,214), behind three OpenAI GPT-Image-2.6 / 2.5 models; 2.7-Flash ranks No. 11 (1,158).
A few notes:
- Editing is the strength here: 2.7 is 64 points above 2.6, and Flash matches or beats some previous-generation flagships. Microsoft says it reaches the frontier in enhancement and restoration, reasoning-based edits, scene and style, and text and symbol editing
- The 4-point lead is within the error bars: the confidence intervals for 1,215 and 1,211 (about ±13) overlap, so the top of the editing board is effectively a tie
- Do not pick a text-to-image model from this board alone: 2.7 trails GPT-Image-2.6 Sunburst by 47 points on T2I. If generation quality is the priority, compare against the GPT-Image-2.6 line first
- Do not mix leaderboard versions: Artificial Analysis changed to the v2.1 methodology in August; older scores are not directly comparable
Key Features
2048×2048 output
Pixel ceiling rises from 2,359,296 to 4,194,304 (+78%), so e-commerce hero images and posters skip an upscale step
Transparent backgrounds
Send
background: "transparent" to get a PNG with an alpha channel — ready for stickers, icons and product cutoutsOne to five reference images
Multi-image fusion in the edits endpoint, referenced as image 1, image 2 and so on; the number of references does not change the price
Precise local edits
Change an object’s color, material or position while keeping the rest of the frame intact, with better consistency across successive edits
Technical Specs
Latency measured on APIYI’s production gateway on November 20, 2026 (1024×1024 unless noted):
Practical Use
Recommended Scenarios
- E-commerce and advertising images: generate at
2048×2048directly; use the two-image edit to place a product into a scene - Transparent assets: stickers, app icons and product cutouts via
background: "transparent" - Brand asset reuse: up to five reference images to combine a logo, a product and a person in one frame
- Posters and signage: wrap Chinese text in quotes — 2.7-Flash rendered shop signs, couplets and chalkboards correctly in testing
Code Example
Best Practices
- Scale timeouts with size: at least 120 s for 1024, at least 180 s for 2048. The image API is synchronous with no task ID — if the client disconnects, the result is lost but the request is still billed
- Use
width+height: the generation endpoint silently ignoressizeand always returns 1024×1024 - Remove three parameters when migrating:
response_format,seedandnegative_promptall return 400 on this series - Upload files for edits: image URLs and base64 JSON are rejected; repeat
image[]for multiple references, neverimage+image2 - Do not select it in chat clients: this series returns 404 on
/v1/chat/completions, so Cherry Studio, LobeChat and similar tools will not work
Pricing and Availability
Pricing
Billed per request, independent of size and reference count: 1024×1024, 1536×1536 and 2048×2048 cost the same, and an edit with one reference image costs the same as one with five. The provider bills by image tokens, so larger images cost more (about $35 per 1,000 images at 1024×1024); APIYI converts that into a flat per-image price so budgets are predictable to the image.
With the current top 20% top-up bonus, the flagship works out to about $0.10 per image and Flash to about $0.05 (bonus tiers depend on the single top-up amount; see the promotion page).
Model prices are aligned with the provider and may change; the table above is for reference only. The Model Pricing tab in the top navigation is authoritative: Model Pricing.
Default and SVIP, both at 1.0x. Any new key can call the models without an application.
Stacking Top-up Bonuses
The per-request prices above stack with APIYI’s top-up bonuses: the top tier is 20% for a single top-up of $3,000 or more, which brings the flagship to about $0.10 per image. See Top-up Promotions for tiers and rules.Conclusion
MAI-Image 2.7 is a meaningful upgrade in editing plus a larger canvas: No. 1 on the editing board, 2048×2048 output, transparent backgrounds — while APIYI’s per-request prices stay at $0.12 / $0.06.- Editing, retouching, local replacement: use 2.7; use 2.7-Flash for volume drafts and variations at half the cost
- Need 2048×2048 or transparent backgrounds: only the 2.7 series supports them; 2.6 remains available but caps at 1536×1536
- Pure text-to-image with maximum quality: 2.7 ranks No. 4 on T2I, behind the GPT-Image-2.6 line — compare before switching
- Already on 2.6: migration is a model-name change; endpoints and parameters are identical
Sources (retrieved November 20, 2026): Microsoft announcement
microsoft.ai/news/introducing-mai-image-2-7/ (November 17, 2026); MAI-Image 2.7 model card and Foundry catalog ai.azure.com/catalog/models/MAI-Image-2.7; Artificial Analysis image editing and text-to-image leaderboards artificialanalysis.ai/image/leaderboard (data as of November 20, 2026); APIYI pricing API and production gateway tests (November 20, 2026; report in code-test/MAI/1120-MAI-Image-2.7-接入测试/).