What it is
MAI-Image-2.6 was announced by Microsoft AI on August 10, 2026, with a smaller MAI-Image-2.6-Flash variant. Microsoft says it improves on MAI-Image-2.5 by 79 Elo on the Arena text-to-image leaderboard, including 91 Elo in text rendering, and launched in second place there.
Both models reached public preview on Microsoft Foundry in September 2026 (model version 2026-07-31) and can be tried in the MAI Playground. They generate and edit images from text and up to five reference images, output PNG up to about 2.4 megapixels, and can optionally ground prompts in Bing search results.
Foundry bills by token: $5 per million text input tokens, $8 per million image input tokens and $38 per million output tokens on Global Standard; Flash costs $1.75, $2.50 and $19. Artificial Analysis converts this to about $39 per 1,000 images at 1024x1024.
Our review
MAI-Image-2.6 is a genuine contender held back by its packaging. Artificial Analysis now ranks it fifth at 1166 Elo, tied with xAI's Grok Imagine Image 2.0 and ten points behind Google's Nano Banana 2.1, with overlapping confidence ranges. Microsoft's launch post cites second place on LMArena's Arena at launch.
On price it sits in the middle of the leaders. Artificial Analysis lists $38.90 per 1,000 images, against $33.60 for Nano Banana 2.1, $60 for Grok Imagine and $210.70 for the top-ranked GPT Image 2.5. The Flash variant is the quieter bargain: 1119 Elo at $19.50 per 1,000, level with MAI-Image-2.5 at under half that model's $48.10.
Foundry's terms will decide most deployments. Both models are in preview, which Microsoft says carries no service-level agreement and is not recommended for production. Default quotas run from 6 requests a minute at tier 1 to 36 at tier 6, with none on the free tier, and each request returns one PNG. Microsoft does not publish how many tokens an image consumes, so budgets rest on Artificial Analysis's conversion.
Two features set it apart from the Google and OpenAI options. Web grounding pulls Bing results into the prompt for real places and current events, at $14 per 1,000 calls, and edits accept up to five reference images. Data falls under Azure's terms for models sold by Azure, which rule out training foundation models on prompts, though Global deployments can process prompts in any region where the model runs.
Where it earns its place
An e-commerce marketing team on Azure producing 50,000 product images a month
Generate lifestyle shots and variants from existing packshots
Send up to five reference photos to the edits endpoint with MAI-Image-2.6-Flash. At Artificial Analysis's $19.50 per 1,000 images, 50,000 a month costs about $975, against about $1,945 on the full model.
A SaaS engineer adding image generation to an Azure-hosted app
Ship an in-product image feature on existing Azure billing and identity
Deploy as Global Standard, authenticate with Entra ID and call /mai/v1/images/generations. Tier 1 allows 6 requests a minute, so a launch needs a quota request, and preview status means no SLA.
A travel publisher making posters of real landmarks
Produce on-brand images of real places with legible text
Turn on web_grounding so the model checks Bing for what a building actually looks like, at $14 per 1,000 calls on top of image cost. Text rendering is the area Microsoft says improved most, by 91 Elo.
Before you commit
- Setup
- Deploy MAI-Image-2.6 (version 2026-07-31) as Global Standard in a Foundry project with the Cognitive Services Contributor role, then POST to the generations or edits endpoint with an API key or Entra ID token.
- Integrations
- REST endpoints at /mai/v1/images/generations and /mai/v1/images/edits on Foundry, the Foundry playground and MAI Playground. Options include auto_aspect_ratio and Bing-backed web_grounding.
- Data and privacy
- Under Azure's terms for models sold by Azure, prompts and outputs are not used to train foundation models or shared with model providers. Abuse monitoring can store flagged content for human review, which managed customers can apply to modify. Microsoft notes that preview models may follow different privacy practices.
- Cost at scale
- Global Standard bills $5 per million text input, $8 per million image input and $38 per million output tokens, about $39 per 1,000 images by Artificial Analysis's count; Flash is roughly half and Data Zone 10% more. 100,000 images a month is about $3,890, or $1,950 on Flash.
- Limits
- Both sides at least 768 pixels and no more than 2,359,296 pixels in total (1536x1536), PNG only, one image per request, 32,000-token prompts in English. Quotas run from 6 to 36 requests a minute; photorealistic edits of minors are blocked by default.
What's good
- Fifth on Artificial Analysis at 1166 Elo, about $39 per 1,000 images
- Flash matches MAI-Image-2.5 quality at $19.50 per 1,000
- Up to five reference images for edits; optional Bing web grounding
- Entra ID auth and Azure data terms
What's not
- Public preview: no SLA, not recommended for production
- Default quota of 6 to 36 requests a minute, one image per request
- English prompts only; Microsoft does not publish tokens per image
Best for
Teams building on Azure that want a strong image model from the same vendor.
Alternatives
MAI-Image-2.6: common questions
- Is MAI-Image-2.6 worth it?
- MAI-Image-2.6 delivers top-five image quality at a fifth of GPT Image 2.5's price, held back by preview terms and quotas that start at six images a minute. We rate it 4.3 out of 5.
- How much does MAI-Image-2.6 cost?
- Foundry Global Standard: $38 per 1M output tokens ($19 Flash), Data Zone 10% more; Artificial Analysis lists about $39 per 1,000 images, $19.50 for Flash.
- What are the downsides of MAI-Image-2.6?
- Public preview: no SLA, not recommended for production; default quota of 6 to 36 requests a minute, one image per request; english prompts only; Microsoft does not publish tokens per image.
- Who is MAI-Image-2.6 best for?
- Teams building on Azure that want a strong image model from the same vendor.
- How much does MAI-Image-2.6 cost for a team or in production?
- Global Standard bills $5 per million text input, $8 per million image input and $38 per million output tokens, about $39 per 1,000 images by Artificial Analysis's count; Flash is roughly half and Data Zone 10% more. 100,000 images a month is about $3,890, or $1,950 on Flash.
- Is MAI-Image-2.6 safe for company data?
- Under Azure's terms for models sold by Azure, prompts and outputs are not used to train foundation models or shared with model providers. Abuse monitoring can store flagged content for human review, which managed customers can apply to modify. Microsoft notes that preview models may follow different privacy practices.
- What are the best MAI-Image-2.6 alternatives?
- The closest alternatives we have reviewed are Nano Banana 2.1, GPT Image 2.5 (Flare and Sunburst) and Seedream 5.0 Pro.
Answers checked Oct 22, 2026.
Make MAI-Image-2.6? Show this rating on your site
