Microsoft MAI-Image-2.5 Review: AI Image Generation Pricing, Quality, and Features
Microsoft just dropped MAI-Image-2.5, and it’s not playing around. Released at Build 2026 on June 2, this is the company’s strongest image model yet — built entirely in-house by the Microsoft AI Superintelligence team under Mustafa Suleyman. It debuted at #2 on Arena’s Image Edit leaderboard and #4 for text-to-image. That puts it ahead of GPT-Image-1.5, Nano Banana Pro 2K, and every FLUX variant. Only GPT-Image-2 and Reve 2.0 sit above it.
I’ve spent the last few days digging into the pricing, benchmarks, features, and real-world output. Here’s everything you need to know.
What Is MAI-Image-2.5?
MAI-Image-2.5 is Microsoft’s flagship text-to-image and image-editing model. It’s not a rebadge of someone else’s model. The MAI (Microsoft AI) lab trained this from scratch on clean, appropriately licensed data — no distillation from third-party models.
The model excels at two things: generating hyper-detailed images from text prompts, and making precise, controllable edits to existing images. Think photorealistic product shots, branding assets, text-heavy designs, and complex scene compositions.
There are two variants shipping together:
- MAI-Image-2.5 — maximum fidelity. Use this for final deliverables where every pixel counts.
- MAI-Image-2.5-Flash — faster and cheaper. Built for high-volume production pipelines where speed matters more than absolute perfection.
Pricing: Surprisingly Aggressive
Here’s where things get interesting. Microsoft is pricing MAI-Image-2.5 aggressively against OpenAI’s suite. All pricing is pay-as-you-go through Microsoft Foundry (formerly Azure AI Foundry), billed per token.
| Model | Text Input (per 1M tokens) | Image Input (per 1M tokens) | Image Output (per 1M tokens) |
|---|---|---|---|
| MAI-Image-2.5 | $5.00 | $8.00 | $47.00 |
| MAI-Image-2.5-Flash | $1.75 | $1.75 | $19.50 |
| MAI-Image-2-Efficient | $5.00 | — | $19.50 |
*Pricing source: Microsoft AI official announcement, June 2, 2026 *
The Flash variant is the real value play here. At $1.75 per 1M input tokens and $19.50 per 1M output tokens, it undercuts most competing models while still delivering production-grade quality. The Efficient variant (MAI-Image-2-Efficient) launched in April 2026 at 41% lower cost than the original MAI-Image-2, with 22% faster render times.
For context, this pricing makes MAI-Image-2.5-Flash one of the most cost-effective high-quality image models available through a major cloud provider. You’re getting Arena top-5 quality at prices that compete with budget-tier options.
Arena Leaderboard Performance
The Arena leaderboard uses blind human preference judging. It’s the closest thing we have to an objective measure of image quality. Here’s where MAI-Image-2.5 lands:
Text-to-Image Rankings (June 3, 2026)
| Rank | Model | Score | Lab |
|---|---|---|---|
| 1 | gpt-image-2 (medium) | 1384 | OpenAI |
| 2 | reve-2.0 | 1280 | Reve |
| 3 | gemini-3.1-flash-image-preview | 1269 | |
| 4 | mai-image-2.5 | 1254 | Microsoft AI |
| 5 | gemini-3-pro-image-preview-2k | 1245 | |
| 6 | gpt-image-1.5-high-fidelity | 1242 | OpenAI |
| 11 | mai-image-2 | 1183 | Microsoft AI |
| 63 | dall-e-3 | 968 | OpenAI |
*Source: Arena.ai Text-to-Image Leaderboard, June 3, 2026 *
Image Edit Rankings (June 3, 2026)
| Rank | Model | Score | Lab |
|---|---|---|---|
| 1 | gpt-image-2 (medium) | 1465 | OpenAI |
| 2 | mai-image-2.5 | 1401 | Microsoft AI |
| 3 | chatgpt-image-latest | 1390 | OpenAI |
| 8 | gpt-image-1.5-high-fidelity | 1373 | OpenAI |
*Source: Arena.ai Image Edit Leaderboard, June 3, 2026 *
MAI-Image-2.5 delivers a +75 ELO point improvement over MAI-Image-2 overall, with the largest gains in Text Rendering (+107 points) and Cartoon, Anime & Fantasy (+90 points). On the editing side, it wins decisively across 12 categories, including image cleanup, background replacement, shadow handling, and text modification.
The model surpassed Nano Banana Pro 2K’s Arena score, which is notable because Gemini’s image models have dominated the upper ranks for months.
Key Features
Photorealism That Holds Up
MAI-Image-2.5 handles lighting, skin tones, and environmental detail with a natural feel that most models struggle to achieve consistently. The model understands scene structure — lighting, scale, spatial relationships — and uses that understanding when making edits.
Example prompts from Microsoft’s model page show a young woman blowing soap bubbles on a rooftop, described with precise detail: “film photography aesthetic, cool desaturated palette with muted tones, broken by vivid pink and green bubble props”. The output captures the specific mood and visual style without hallucinating details.
Precision Editing at Every Pixel
This is where MAI-Image-2.5 really shines. The model supports fine-grained, localized edits without changing the rest of the image. Microsoft’s demos show:
- Changing a tote bag color from beige to orange while preserving lighting and shadows
- Adding peonies to an existing image with correct perspective
- Updating text on a product label (“Vanilla Watermelon” to “Licorice Auckland”) while the background composition stays intact
- Removing motion blur or cleaning up backgrounds
The model preserves facial identity across edits, which is a notoriously hard challenge. Change the pose, expression, or viewpoint, and the person still looks like the same person.
Design-Aware Generation
MAI-Image-2.5 was built with commercial design work in mind. It handles branding mockups, product shots, and packaging design remarkably well. The model card demo shows a fictional brand (“ORVA”) consistently rendered across candle jars, skincare bottles, and lifestyle product shots with coherent branding.
In-Image Text Rendering
Text in AI-generated images is usually a mess. MAI-Image-2.5 is a noticeable step forward. The +107 ELO gain in Text Rendering over MAI-Image-2 reflects this improvement. Product labels, posters, and typographic layouts come out clean and readable.
WPP’s Global Chief Creative Officer Rob Reilly called MAI-Image-2 “a genuine game-changer” that “deeply respects the sheer craft involved in generating real-world, campaign-ready images”. Shutterstock’s Principal Product Manager noted “strong progress in prompt fidelity and creative usability across a range of workflows”.
Comparison With the Competition
vs OpenAI (GPT-Image / DALL-E)
GPT-Image-2 still holds the #1 spot on both Arena leaderboards, but MAI-Image-2.5 is closer to it on image editing (1401 vs 1465) than any other model. On text-to-image, the gap is wider (1254 vs 1384), but MAI-Image-2.5 already beats GPT-Image-1.5 High Fidelity (1242) and matches or exceeds GPT-Image-1.5 across most categories.
The price-to-performance ratio favors Microsoft. While OpenAI doesn’t publicly list GPT-Image-2 pricing in the same token-based structure, MAI-Image-2.5-Flash at $19.50/1M output tokens is clearly positioned as the budget-friendly alternative that doesn’t sacrifice much quality.
DALL-E 3, which was OpenAI’s flagship image model a generation ago, now sits at rank 63 on the Arena with a score of 968. MAI-Image-2.5’s 1254 score makes that comparison almost unfair.
vs Google (Gemini / Nano Banana)
Google’s Gemini 3.1 Flash and Pro image models are MAI-Image-2.5’s closest competitors in the ranking spread. MAI-Image-2.5 (1254) narrowly edges out Gemini 3 Pro Image Preview 2K (1245) on text-to-image. On image editing, the Microsoft model pulls ahead more decisively (1401 vs 1388 for Gemini 3 Pro Image Preview 2K).
Microsoft explicitly states that MAI-Image-2.5 “surpasses the Arena score of Nano Banana Pro”, which covers both the Flash and Pro Google variants.
vs Midjourney
Midjourney isn’t on the Arena leaderboard, so direct numerical comparison isn’t possible through that channel. However, Midjourney’s strengths — photorealistic scenes, artistic styling, strong community workflows — are areas where MAI-Image-2.5 competes head-on. The Microsoft model’s emphasis on photorealism, design-ready output, and text rendering addresses use cases where Midjourney has traditionally been the go-to.
The key differentiator is integration. Midjourney operates primarily through Discord. MAI-Image-2.5 is available through an API on Microsoft Foundry, integrated into PowerPoint and OneDrive, and accessible via OpenRouter, Fireworks, and Baseten. For enterprise teams building products on top of image generation, the API-first approach matters.
vs Stable Diffusion / FLUX
Stable Diffusion 3.5 Large scores 938 on the Arena (rank 65), and FLUX.2 Max scores 1163 (rank 16). MAI-Image-2.5 at 1254 is a full tier above both.
That said, Stable Diffusion and FLUX offer something MAI-Image-2.5 doesn’t: open weights and self-hosted deployment. If you need air-gapped, on-premises image generation, MAI-Image-2.5 isn’t your answer. But if you’re building on Azure and want managed, serverless image AI, the quality gap is substantial.
vs Adobe Firefly
Adobe Firefly isn’t on the Arena leaderboard as a standalone model. Its strength is integration with Adobe’s Creative Cloud suite. MAI-Image-2.5 competes on a different axis — it’s a raw API model rather than a design-tool-embedded feature. For developers building image generation into apps, MAI-Image-2.5 offers more flexibility. For designers working inside Photoshop, Firefly’s deep integration remains valuable.
The MAI Image Model Family
MAI-Image-2.5 is the latest in a rapid release cadence from Microsoft’s lab:
- October 2025 — MAI-Image-1 debuts in the top 10 on LMArena
- March 19, 2026 — MAI-Image-2 launches as the #3 text-to-image model family
- April 14, 2026 — MAI-Image-2-Efficient: flagship quality at 41% lower cost
- June 2, 2026 — MAI-Image-2.5 and MAI-Image-2.5-Flash launch
That’s three major releases in under three months. The pace suggests Microsoft is running a serious, well-funded AI lab with iterative improvement cycles. Mustafa Suleyman described their approach as building “a hill-climbing machine” — an organization that continuously improves cycle after cycle.
Where to Use It
Best Use Cases
- E-commerce product photography — Generate clean, brand-consistent product shots at scale
- Marketing creative iterations — Rapidly produce ad variants with controlled edits
- Branding and packaging design — Prototype logos, labels, and packaging concepts
- Presentation design — Direct integration with PowerPoint for slide-ready visuals
- Photo editing and cleanup — OneDrive integration for removing distractions, changing backgrounds
Where to Be Cautious
- Identity verification or legal contexts — Generated images can contain plausible but inaccurate details
- Medical or financial imagery — Not validated for these domains
- News photography — Generated images should not be presented as real photographs
- Air-gapped deployments — No self-hosted / on-premises option (serverless only)
Safety and Data Practices
Microsoft emphasizes that MAI-Image-2.5 was trained on “clean, enterprise-grade data lineage” without relying on “unlicensed or opaque data”. The model includes layered safety guardrails with prompt and output filtering to block harmful content.
Like all generative models, MAI-Image-2.5 can reflect biases present in its training data. Microsoft recommends reviewing outputs before using them in sensitive contexts.
On the data privacy side, Microsoft says your prompts and generated images are not used to train or improve their models. This is the standard Azure AI Services data handling policy.
The Bottom Line
MAI-Image-2.5 is the strongest image model Microsoft has ever shipped, and it earns its spot in the Arena top 5. The editing capabilities in particular are a standout — being #2 on the Image Edit leaderboard, ahead of every Google and xAI model, is no small feat.
The pricing makes it genuinely competitive. At $5/1M text input and $47/1M image output for the flagship, or $1.75/$19.50 for Flash, businesses can actually budget for production-scale image generation without a second mortgage.
If you’re already on Azure, there’s no reason not to try it. If you’re shopping for an AI image API and want to avoid the Discord-bot workflow of Midjourney or the higher prices of OpenAI’s top-tier models, MAI-Image-2.5 deserves a serious look.
The MAI lab has gone from zero to top-5 in under a year. At this velocity, the gap between MAI-Image and the #1 spot might not last long.
Sources
- Microsoft AI, “Building a hill-climbing machine: Launching seven new MAI models,” June 2, 2026. https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/
- Arena.ai, “Text-to-Image Leaderboard,” accessed June 3, 2026. https://arena.ai/leaderboard/text-to-image
- Mustafa Suleyman, “Building a hill-climbing machine: Launching seven new MAI models,” Microsoft AI Blog, June 2, 2026. https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/
- Microsoft AI Superintelligence Team, “MAI-Image-2.5 launches at No. 2 for image editing on Arena,” June 2, 2026. https://microsoft.ai/news/introducing-mai-image-2-5/
- Microsoft AI, “MAI-Image-2-Efficient: Flagship Quality, 41% Lower Cost,” April 14, 2026. https://microsoft.ai/news/mai-image-2-efficient/
- Arena.ai, “Image Edit Leaderboard,” accessed June 3, 2026. https://arena.ai/leaderboard/image-edit
- Microsoft AI, “MAI-Image-2.5 Model Page,” accessed June 5, 2026. https://microsoft.ai/models/mai-image-2-5/
- Microsoft AI, “Introducing MAI-Image-1, debuting in the top 10 on LMArena,” October 2025. https://microsoft.ai/news/introducing-mai-image-1-debuting-in-the-top-10-on-lmarena/
- Microsoft AI, “Introducing MAI-Image-2: for limitless creativity,” March 19, 2026. https://microsoft.ai/news/introducing-mai-image-2/
- Microsoft Azure, “Azure AI Foundry Models Pricing,” accessed June 5, 2026. https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/microsoft/
Disclaimer: Arena scores are preliminary as of June 3, 2026 and may shift as more votes accumulate. Pricing is subject to change. This review is not sponsored by Microsoft.