MAI-Image-2.6 is Microsoft AI's image model, released on 4 September 2026, and Microsoft reports it gained 79 Elo over MAI-Image-2.5 on Arena text-to-image. It targets teams that edit from up to 5 reference images at up to 1.5K resolution and that want a cheaper Flash sibling for high-volume work.
Can you use it?
- OpenRouter: yes (Pay per use through OpenRouter).
- MAI Playground: yes. Named as a place to try the model in Microsoft's launch post; plans not stated.
- Free plan: no.
Checked 10 Nov against Microsoft AI's own pages. Compare access across frontier models
What changed
MAI-Image-2.6 is a diffusion-based text-to-image and image-editing model from Microsoft AI, sold through Microsoft Foundry. Artificial Analysis rated it 1168 Elo for text-to-image and 1149 for editing on 10 November 2026. It takes text and images as input, returns images, and has a 32,000-token context window.
Provider: Microsoft AI · Family: MAI-Image
More about Microsoft AI on Hokai
Context window: 32,000 tokens · Max output: 4,096
Input modalities: text, image · Output: image
About MAI-Image-2.6
MAI-Image-2.6 is a diffusion-based text-to-image and image-editing model from Microsoft AI, part of Microsoft. Microsoft's Foundry model card says it transforms random noise into an image using a flow-matching loss, and describes it for both text-to-image synthesis and controlled image-to-image editing. Microsoft announced it on 4 September 2026 as the precision tier of a two-model family, next to the faster MAI-Image-2.6-Flash, and it follows MAI-Image-2.5.
On independent boards, Artificial Analysis listed MAI-Image-2.6 at 1168 Elo for text-to-image and 1149 Elo for image editing when Hokai read its image leaderboard on 10 November 2026. Microsoft itself reports that the model ranked No. 2 on the Arena text-to-image leaderboard, gained 79 Elo over MAI-Image-2.5 overall, and gained 91 Elo on text rendering (vendor-reported, announcement of 27 October 2026). Those are Microsoft's claims, not a Hokai measurement.
The Foundry model card lists a 32,000-token context window and a 4,096-token output limit, with text and images as input and images as output. Microsoft's launch post adds multi-reference editing, web grounding, dynamic aspect ratios and resolutions up to 1.5K. OpenRouter's listing says the multi-reference mode combines people, products, styles and scenes from up to five input images.
The model is sold through Microsoft Foundry and is also listed on OpenRouter. The Foundry card shows its lifecycle as Preview and its version as 2026-07-31, while Microsoft's 27 October 2026 update says the model is generally available in Foundry. Check the card before you commit a production workload. Microsoft has not published a 2.6 line on its Foundry pricing page, so the price section below reports only what third parties list.
Microsoft AI has since released MAI Image 2.7, an editing-focused model that launched in preview on 27 October 2026 with text-to-image generation promised later. It does not replace 2.6 for generating images from text.
Who should not use it: on the Artificial Analysis board read on 10 November 2026, GPT Image 2.6 in its Sunburst (max) setting scored 1259 Elo for text-to-image and 1208 for editing, ahead of MAI-Image-2.6 on both. Teams that want the top score on either board should weigh it first. Teams already on Google's stack may prefer Gemini 3 Pro Image, and teams that want a different model family can look at FLUX 3. Anyone who needs published weights, local inference or a published data-retention policy will not find those for this model in the sources Hokai read.
Pricing
Microsoft has not listed MAI-Image-2.6 on its Foundry pricing page (read 10 November 2026), so Hokai shows no official rate. OpenRouter lists $8 per 1M input tokens and $38 per 1M output tokens. Artificial Analysis prices it at $38.91 per 1,000 images, and its Flash sibling at $19.46 per 1,000 images.
Key Features
- Multi-reference editing: Blends subjects, brand products and visual styles from several uploaded references into one edit.
- Web grounding: Pulls current information from the web into the image, per Microsoft's launch post.
- Dynamic aspect ratios: Lets the model choose the composition format, with resolution up to 1.5K.
Pros
- Places in the 1160s Elo range for text-to-image on an independent leaderboard.
- A cheaper Flash sibling costs half as much per 1,000 images, by Artificial Analysis pricing.
Cons
- Trails GPT Image 2.6 Sunburst (max) on both independent leaderboards.
- No official per-token price on Microsoft's Foundry pricing page at the time of reading.
Frequently Asked Questions
How much does MAI-Image-2.6 cost?
Microsoft has not listed MAI-Image-2.6 on its Foundry pricing page as of 10 November 2026, so there is no official rate to quote. OpenRouter lists $8 per 1M input tokens and $38 per 1M output tokens. Artificial Analysis prices it at $38.91 per 1,000 images, and the Flash version at $19.46 per 1,000 images.
Is MAI-Image-2.6 open source or free to use?
No. MAI-Image-2.6 is proprietary and Microsoft publishes no weights. You reach it through Microsoft Foundry or OpenRouter, and Microsoft's launch post also points to MAI Playground for trying it.
What can MAI-Image-2.6 take as input and return?
The Foundry model card lists text and images as input and images as output, with a 32,000-token context window and a 4,096-token output limit. Microsoft says it supports multi-reference editing, web grounding, dynamic aspect ratios and resolution up to 1.5K.
MAI-Image-2.6 vs GPT Image 2.6: which scores higher?
On the Artificial Analysis board read on 10 November 2026, GPT Image 2.6 Sunburst (max) scored 1259 Elo for text-to-image and 1208 for editing. MAI-Image-2.6 scored 1168 and 1149, so GPT Image 2.6 leads on both axes there.
Is MAI-Image-2.6 generally available?
Microsoft's 27 October 2026 update says MAI Image 2.6 and 2.6 Flash are generally available in Microsoft Foundry. The Foundry model card still showed the lifecycle as Preview on 10 November 2026, so confirm the status in Foundry before a production launch.
Top Alternatives
- GPT Image 2.6: Pick GPT Image 2.6 for the higher Artificial Analysis scores; pick MAI-Image-2.6 for Foundry hosting and multi-reference editing.