Chrome Extension
Microsoft quietly built one of the strongest image generation models available right now. MAI-Image-2.5 debuted at number three on Arena's text-to-image leaderboard on May 26, 2026, and by June 2, it had climbed to number two on Arena's image editing leaderboard. I have been testing it across commercial, creative, editorial, and UI design workflows, and the results are consistent enough that I want to document every prompting pattern that works.
This guide covers the full specs, both pricing tiers, the four core capabilities, and twenty-plus ready-to-use prompts organized by use case. If you have been using GPT-image-2, Nano Banana or Midjourney and have not tested MAI-Image-2.5 yet, this is the article that will change your evaluation.
Checkout our free Mai-Image Prompt Generator here.
Key Takeaways
What Is MAI-Image-2.5? The Specs Worth Knowing
MAI-Image-2.5 is a text-to-image and image-editing model built by Microsoft's Superintelligence team. It is part of the MAI model family, which also includes reasoning, coding, transcription, and voice models. The full stack was announced at Microsoft Build 2026 on June 2, 2026. You can read the official launch announcement here.
| Specification | Detail |
|---|---|
| Release date | June 2, 2026 (Arena debut May 26, 2026) |
| Developer | Microsoft MAI Superintelligence Team |
| Arena text-to-image rank | Number 3 (score 1,254) |
| Arena image edit rank | Number 2 (score 1,401) |
| Improvement over MAI-Image-2 | +75 overall Arena points |
| Text Rendering improvement | +107 Arena points |
| Cartoon, Anime and Fantasy | +90 Arena points |
| Variants | MAI-Image-2.5 (standard) and MAI-Image-2.5-Flash |
| Input types | Text prompts, image uploads |
| Output | Generated or edited images |
| Integration | Azure AI Foundry, MAI Playground, OpenRouter, PowerPoint, OneDrive |
The model understands scene structure, lighting, scale, and spatial relationships. When you ask it to add an object to a scene, it places that object with the correct perspective and shadows relative to the existing elements. That visual reasoning capability is what separates it from earlier MAI-Image versions and from most mid-tier generation models.
Pricing Breakdown: Standard vs Flash
MAI-Image-2.5 ships in two pricing tiers designed for different production workloads.
| Pricing Item | MAI-Image-2.5 (Standard) | MAI-Image-2.5-Flash |
|---|---|---|
| Text input | $5 per 1M tokens | $1.75 per 1M tokens |
| Image input | $8 per 1M tokens | $1.75 per 1M tokens |
| Image output | $47 per 1M tokens | $19.50 per 1M tokens |
The Flash variant is built for high-volume, latency-sensitive production workflows where speed and cost matter more than maximum fidelity. My recommendation is to prototype with the standard variant to validate your prompt structure, then switch to Flash for bulk generation once you have the output quality dialed in.
For context on where this sits in the market: DALL-E 3 via Azure runs approximately $40 per million output tokens. MAI-Image-2.5-Flash at $19.50 is more than 50 percent cheaper. The standard tier at $47 is slightly above DALL-E 3, but the Arena leaderboard position reflects the quality difference. You are paying a premium for better prompt adherence, sharper text rendering, and image editing capability.
For image editing workflows that submit an existing image plus a modification instruction, the image input token cost of $8 per million on standard (or $1.75 on Flash) is charged on top of the output cost.
Arena Rankings: What the Numbers Actually Mean
Arena's leaderboard uses blind human preference voting. Two model outputs are shown side by side without labels, and human judges vote for the one they prefer. The final ELO score reflects aggregate preferences across thousands of comparisons.
MAI-Image-2.5 holds a score of 1,254 on text-to-image, placing it at number three globally. Above it are OpenAI's GPT-Image-2 and Google's top-ranked image model. The model beat GPT-Image-1.5 and Google's Nano Banana Pro 2K.
On the image editing leaderboard, MAI-Image-2.5 holds a score of 1,401 with a margin of plus or minus 8 points, placing it at number two. Only GPT-Image-2 (medium) ranks higher.
The category-level breakdown from Arena shows where MAI-Image-2.5 wins most reliably:
This tells you exactly which use cases to prioritize when testing the model. If your workflow involves any of these categories, MAI-Image-2.5 is worth a direct comparison against whatever you currently use.
The 4 Core Capabilities You Need to Understand
1. Text-to-Image Generation
The foundation of the model. You provide a text prompt and receive a generated image. MAI-Image-2.5 handles complex, multi-element prompts better than its predecessors. When you specify multiple subjects, spatial arrangements, style references, and lighting conditions simultaneously, the model honors more of those constraints at the same time.
2. Image Editing with Upload
This is the major new capability in 2.5. You submit an existing image alongside a text instruction. The model makes the specified change while preserving the rest of the image. This is what enables localized edits: change a background, replace an object, update text on a sign, remove a distraction, adjust lighting in one area, all without touching the parts of the image you want to keep.
3. Text Rendering Within Images
Words embedded in generated images have historically been a weak point for AI image models. MAI-Image-2.5 closes this gap significantly. Posters, labels, packaging, presentation slides, signage, and branded assets with visible text are now practical outputs. The +107 Arena point improvement in text rendering is the most dramatic single-category gain in the 2.5 release.
4. Face and Identity Consistency
When you edit an image that contains a face, MAI-Image-2.5 preserves the facial identity across the edit. The same person's likeness is maintained even through changes in pose, expression, background, or viewpoint. This is critical for brand photography workflows where a subject's consistent appearance across a series of assets is required.
20+ Prompts for Every Use Case
I used Microsoft's Playground tool to generate the following images.
Commercial Product Photography
Use case 1: Clean product shot with white background
A luxury skincare serum bottle, glass with a matte gold cap, photographed against a pure white background. Studio lighting with a soft key light from the upper left and a subtle rim light behind. No shadows on the background. The label reads "LUMÉ RADIANCE SERUM" in a clean serif font. Editorial product photography style. 4K resolution.
Use case 2: Lifestyle product placement
A matte black travel coffee thermos sitting on a weathered wooden table outdoors at a campsite. Morning light filtering through pine trees. A hand is reaching into frame from the right to pick up the thermos. The label on the thermos reads "BREW & GO" in bold white sans-serif type. Warm amber color grade. Commercial lifestyle photography.
Use case 3: Packaging flat lay
Top-down flat lay of a skincare gift set: a small glass jar, a roller bottle, and a cotton drawstring pouch, all in soft beige tones. Arranged on a smooth marble surface with dried lavender sprigs between the products. Labels on each product read "CALM," "GLOW," and "RESTORE" respectively. Soft diffused daylight. Commercial packaging photography.
Use case 4: Before and after product edit (image editing)
[Upload: existing product photo with cluttered background]
Replace the background with a solid pale sage green. Keep the product, lighting, and shadows exactly as they are. Do not change the product or any text on its label.
Social Media and Marketing Assets
Use case 5: Instagram story graphic with text
Vertical social media story graphic, 9:16 ratio. Deep navy blue background with a subtle gradient to midnight blue at the bottom. Bold white headline text centered at the top: "SALE ENDS TONIGHT". Below it in smaller gold text: "Up to 60% off everything". A thin horizontal gold divider line between the two text blocks. Minimal, high-end retail aesthetic. No illustrations, text only.
Use case 6: LinkedIn banner for a tech company
Wide horizontal banner, LinkedIn dimensions. Dark charcoal background with a subtle blue-to-purple gradient from left to right. On the left side: company logo placeholder as a glowing white geometric mark. Center text in clean white sans-serif: "Building the Future of Work" on the first line, "AI-Powered HR Solutions" on the second line in lighter weight. Right side: abstract network visualization in pale blue lines. Corporate technology aesthetic.
Use case 7: YouTube thumbnail with text overlay
YouTube thumbnail, high contrast. A person in a white hoodie sitting at a desk with multiple monitors, face lit by screen glow. Bold yellow text in the upper left: "I TESTED". Large white text centered: "MAI-Image-2.5". Bold red text bottom right: "RESULTS". Dark background with vignette. Maximum contrast for small-screen readability.
Use case 8: Facebook ad creative for a fitness brand
Square social ad, 1:1 ratio. Split composition: left half shows a protein powder canister with "PEAK FORM WHEY" label prominently displayed. Right half shows a dark background with white headline text: "30g Protein Per Serving" and subtext: "Now in 3 new flavors." Orange accent color for the subtext. Gym photography aesthetic.
Poster and Print Design
Use case 9: Event poster with full type layout
A concert poster for a jazz festival. Dark burgundy background with textured paper grain overlay. Large stylized title text at the top: "MOONRISE JAZZ FESTIVAL" in an art deco serif font, gold color. Center illustration: a silhouette of a saxophonist in blue light. Below the illustration in clean white text: "June 28 and 29, 2026". "Riverside Amphitheatre, Austin TX" in smaller text. Ticket purchase URL at the bottom: "moonrisejazz.com". Vintage poster aesthetic.
Use case 10: Real estate listing poster
Real estate marketing flyer, portrait orientation. Hero image: an exterior view of a modern white-rendered house with floor-to-ceiling windows and a slate grey front door, photographed at golden hour. Below the image in a clean grey box: large bold text "4 BED | 3 BATH | 2,400 SQ FT" and below it "Listed at $1.2M". At the bottom: agent name placeholder, phone number, and small agency logo space. High-end real estate marketing aesthetic.
Use case 11: Restaurant menu cover
A restaurant menu cover in print quality. Dark forest green linen texture background. Gold foil-effect centered text: the word "VERDURE" in an elegant italic serif at the top, and below it "Farm to Table Dining" in a lighter weight. A single botanical illustration of rosemary in fine line gold below the text. No other elements. Luxury print design aesthetic.
Presentation and Slide Assets
Use case 12: PowerPoint slide hero image
Wide horizontal presentation slide background, 16:9 ratio. Clean white background with a bold dark blue geometric shape filling the left 40 percent. On the blue shape: white text reading "Q3 2026" on the first line and "Performance Review" on the second line in a lighter weight. Right side: abstract upward trending graph lines in a light blue-to-turquoise gradient on the white area. Professional corporate presentation aesthetic.
Use case 13: Data visualization illustration for a slide
A flat design illustration of a city skyline made from bar chart shapes. Each building is a different height representing different data values. The tallest bar is labeled "Revenue" in white text, medium bars labeled "Engagement" and "Retention". Bright color palette: coral, teal, and gold. White background. Infographic style. No grid, no axes, stylized visualization only.
UI and App Design Mockups
Use case 14: Mobile app onboarding screen
A mobile app onboarding screen mockup, iPhone frame. Dark mode UI. Screen content: a clean centered illustration of a glowing target with an arrow in the center, in blue-to-purple gradient. Below the illustration: bold white headline "Track Every Goal" and subtext in light grey "Your progress. Your pace. No distractions." At the bottom: a primary blue CTA button with white text "Get Started". Clean iOS design aesthetic.
Use case 15: SaaS dashboard preview illustration
A product dashboard UI illustration for a project management SaaS. Shows a dark-mode browser window with sidebar navigation on the left and a main content area with a kanban board: three columns labeled "To Do", "In Progress", and "Done" with card stacks in each. Status pills in green, yellow, and red. Top navigation bar with a search field and user avatar. Modern SaaS design aesthetic. Flat illustration style, no real data.
Creative and Artistic Output
Use case 16: Cinematic scene illustration
A cinematic wide-angle illustration of a lone astronaut standing on a red sand dune on Mars. The astronaut faces away from the viewer, looking toward a massive rust-colored sky with two distant moons visible. Long shadows across the dunes. The suit is white with an orange visor. Atmosphere: vast, isolated, awe-inspiring. Painterly style with photorealistic lighting. Widescreen 16:9.
Use case 17: Character portrait with consistency requirement
A portrait illustration of a young woman with short dark hair, olive skin, and sharp cheekbones. She is wearing a black turtleneck. Expression: confident, slightly amused. Lighting: a single soft key light from the upper left casting a gentle shadow on the right side of her face. Style: editorial illustration, semi-realistic. Clean white background.
Use case 18: Anime-style scene
A Studio Ghibli-inspired illustration of a girl riding a bicycle through a rain-soaked Japanese countryside road. Rice fields on both sides, a single sakura tree in full bloom to the right. Grey sky with soft light diffusing through clouds. The girl is wearing a yellow raincoat and has a green school bag. Warm and slightly melancholic mood. Painterly anime style.
Use case 19: Architecture concept visualization
An architectural visualization of a minimalist house in a forest setting. The structure is a single-story volume with floor-to-ceiling glazing, dark weathered timber cladding, and a flat roof with a narrow overhang. Surrounded by tall birch trees. Blue hour lighting, interior lights visible through the glass. Shot angle: slightly elevated three-quarter view. Photorealistic rendering style.
Brand and Logo Asset Generation
Use case 20: Logo lockup on brand background
A brand identity mockup. Centered on a deep navy blue background: a geometric logo mark consisting of two overlapping hexagons in electric yellow. Below the mark, the company name "NEXVAULT" in bold white uppercase sans-serif. Below the name in lighter weight: "Secure Cloud Infrastructure". Clean, corporate technology brand aesthetic. No gradients except the overlap area of the hexagons.
Use case 21: Business card design mockup
A premium business card mockup, horizontal orientation. Front side: matte black background. White text in the upper left: "AIKO TANAKA" in bold, and below it "Creative Director" in light weight. Lower right: "aiko@studioform.co" and "+1 415 222 0011" in small text. A single thin white horizontal rule near the bottom. Ultra-clean, luxury branding aesthetic. No logo, typography only.
Editing and Retouching Workflows
Use case 22: Background removal and replacement
[Upload: outdoor photo with a distracting parking lot background] Remove the background entirely and replace it with a soft out-of-focus bokeh of autumn leaves in warm amber and orange tones. Keep the subject, their clothing, hair, and lighting exactly as they appear. Maintain the edge detail around the hair.
Use case 23: Object removal from an existing image
[Upload: product photo with a small price tag sticker visible] Remove the price tag sticker in the lower left corner of the product. Fill in the area with the underlying surface texture. Do not change anything else in the image.
Use case 24: Style transfer with identity preservation
[Upload: professional headshot photo] Transform the background to look like a modern office environment with soft bokeh. Keep the subject's face, hair color, skin tone, clothing, and expression exactly as they are. Adjust the lighting slightly to match the new background.
Educational and Explainer Content
Use case 25: Infographic-style diagram
A clean infographic diagram showing the water cycle. Three stages labeled with large bold text: "EVAPORATION" at the bottom left with upward arrows from a blue ocean surface, "CONDENSATION" at the top center with a stylized cloud, and "PRECIPITATION" at the right with falling rain arrows returning to the ocean. Simple, bright colors: blue for water, white for clouds, light blue-green for the land mass at the right edge. Flat vector style on a white background. Educational illustration.
Prompting Best Practices for MAI-Image-2.5
After testing these use cases, I found consistent patterns that improve output quality across all categories.
Always specify the visual context with precision. MAI-Image-2.5 responds to specific descriptors better than vague ones. "A luxury skincare bottle" produces less consistent results than "a 100ml glass serum bottle with a brushed gold cap and a minimal white label." The more specific you are about shape, material, and finish, the more reliable the output.
Anchor text placement explicitly. For any use case involving text within an image, tell the model exactly where the text goes, what size it is relative to the composition, and the font weight. "Bold white uppercase sans-serif text centered at the top" outperforms "text saying X at the top." The text rendering improvement in 2.5 is real, but it still benefits from clear placement instructions.
Name the photography or illustration style. Terms like "editorial product photography," "architectural visualization," "flat vector illustration," "Studio Ghibli-inspired," and "cinematic wide-angle" signal the visual register to the model. These are not decorative additions. They change the output category.
Specify lighting. Lighting descriptions such as "golden hour," "soft diffused daylight," "single key light from upper left," and "screen glow" are among the highest-leverage additions to any prompt. Light defines the mood and spatial quality of an image more than almost any other single parameter.
Use the preserve instruction for editing prompts. When submitting an existing image for editing, always end the instruction with what you want left unchanged. "Remove the background. Keep the product, its label text, and the drop shadow exactly as they are." Without this, the model may interpret the edit as permission to adjust adjacent elements.
Layer style and technical instructions together. The strongest prompts combine style direction ("luxury print design aesthetic"), technical specs ("portrait orientation," "4K resolution"), and content details ("gold foil text," "single botanical illustration") in the same prompt. Separating these into weak descriptors produces weaker outputs.
MAI-Image-2.5 vs the Competition
| Model | Arena Text-to-Image Rank | Image Editing | Text Rendering | Starting Output Price |
|---|---|---|---|---|
| MAI-Image-2.5 | Number 3 | Yes | Strong (+107 vs prior) | $47/M output tokens |
| GPT-Image-2 | Number 1 | Yes | Strong | Comparable |
| Google Nano Banana 2 | Number 2 | Yes | Strong | Comparable |
| MAI-Image-2.5-Flash | Not ranked separately | Yes | Strong | $19.50/M output tokens |
| DALL-E 3 (Azure) | Below top 3 | Limited | Moderate | ~$40/M output tokens |
| Midjourney v8 | Below top 3 | Limited | Moderate | Subscription model |
The competitive position for MAI-Image-2.5 is clearest in three scenarios. First, if you are already building inside Azure and Microsoft 365, using MAI-Image-2.5 avoids a third-party vendor contract entirely. Second, if text rendering within images is a requirement for your workflow, the +107 Arena improvement makes 2.5 the practical choice among models in this price range. Third, if you need image editing at Flash pricing, $19.50 per million output tokens is one of the most cost-effective editing-capable image models currently available.
Where GPT-Image-2 likely maintains an edge is on the most photorealistic portrait and human figure outputs. OpenAI has trained heavily on this category and the top Arena position reflects it. For editorial human photography at maximum fidelity, testing both models on your specific use case is worth the time before committing to a pipeline.
How to Access MAI-Image-2.5
MAI Playground. The fastest way to test the model with no setup required. Go to playground.microsoft.ai/chat, select MAI-Image-2.5 from the model list, and start generating. Both standard and Flash variants are available. No Azure subscription required for initial testing.
Azure AI Foundry. For production workloads. Access through the Foundry model card. API documentation is at learn.microsoft.com/azure/foundry/foundry-models/how-to/use-foundry-models-mai. Full API access requires an Azure account and Foundry setup.
OpenRouter. If you are already routing model calls through OpenRouter, MAI-Image-2.5 is available at openrouter.ai/microsoft/mai-image-2.5. This is the fastest path for developers who have existing OpenRouter integrations and want to add image generation without additional authentication setup.
PowerPoint Copilot. If you have a Microsoft 365 subscription with Copilot enabled, MAI-Image-2.5 is live inside PowerPoint for slide and presentation image generation. See Microsoft's support documentation for setup details.
OneDrive. Rolling out to OneDrive users for photo editing workflows. Details at Microsoft's OneDrive blog.
Frequently Asked Questions (FAQs)
What is MAI-Image-2.5 and who made it?
MAI-Image-2.5 is Microsoft's latest text-to-image and image-editing model, built by the Microsoft Superintelligence team and released on June 2, 2026. It ranks number three on Arena's text-to-image leaderboard and number two on the image editing leaderboard as of June 2026. It is available through Azure AI Foundry, the MAI Playground, OpenRouter, and embedded in Microsoft 365 products.
What is the difference between MAI-Image-2.5 and MAI-Image-2.5-Flash?
MAI-Image-2.5 (standard) is optimized for maximum image quality at $47 per million output tokens. MAI-Image-2.5-Flash is a faster, cheaper variant at $19.50 per million output tokens. Both support text-to-image generation and image editing with uploaded photos. Use the standard variant for final production assets and the Flash variant for high-volume generation workflows where speed and cost matter more than peak fidelity.
Can MAI-Image-2.5 edit existing images?
Yes. This is the major new capability in 2.5 compared to MAI-Image-2. You upload an existing image alongside a text instruction and the model makes the specified change while preserving the rest of the image. Supported editing types include background replacement, object removal, text updates, lighting adjustments, and style transfers.
How good is the text rendering in MAI-Image-2.5?
Text rendering improved by +107 Arena points over MAI-Image-2, making it one of the largest single-category gains in the 2.5 release. The model handles short text strings in posters, labels, packaging, business cards, and slide assets reliably. For longer text blocks or highly stylized typography, providing explicit placement and style instructions in your prompt consistently improves results.
Where does MAI-Image-2.5 rank compared to GPT-Image-2 and Midjourney?
MAI-Image-2.5 ranks number three on Arena's text-to-image leaderboard and number two on image editing, placing it above DALL-E 3, Midjourney v8, and most open-source alternatives. GPT-Image-2 holds the number one position on both leaderboards. Midjourney is stronger for highly artistic and stylized outputs but lacks native image editing capability and does not embed into enterprise software pipelines the way MAI-Image-2.5 does through Microsoft 365.
Is MAI-Image-2.5 available without an Azure subscription?
Yes. The MAI Playground at playground.microsoft.ai/chat lets you test both standard and Flash variants without an Azure subscription. For production API access, Azure AI Foundry requires an Azure account. OpenRouter provides API access through their developer platform for builders who prefer that routing.
Final Thoughts
MAI-Image-2.5 earns its Arena ranking. The text rendering improvement alone makes it worth testing for anyone who has avoided AI image generation because of garbled type. The image editing capability closes the gap with GPT-Image-2 on a category that was missing entirely from MAI-Image-2. And the Flash pricing at $19.50 per million output tokens puts production-grade edited image generation within reach for use cases that would previously have been too expensive to automate.
The twenty-five prompts in this guide are starting points, not finished recipes. Test each one against your actual content needs, adjust the specificity of your material descriptions and lighting instructions, and add the preserve instruction on any editing prompt where you have elements you cannot afford to lose.
MAI-Image-2.5 is the strongest image model inside the Microsoft ecosystem and a legitimate competitor to the top-ranked alternatives outside it. For developers building in Azure, the absence of a third-party vendor contract is worth something beyond the benchmark score. Start with the MAI Playground, validate on your use case, and scale through Azure AI Foundry.
