What Is MAI-Image-2.6: $0.039 Per Image

What Is MAI-Image-2.6? No. 2 on Arena at $0.039 Per Image

MAI-Image-2.6 is Microsoft's generative image model for creating and editing images from text, source images, multiple visual references, and grounded web information. It is available in public preview through Microsoft Foundry, with access also announced for MAI Playground and Arena. Reported benchmark data places the flagship at No. 2 for text-to-image and around $0.039 per output image, while its Flash variant targets lower latency and cost.

Key Takeaways

• MAI-Image-2.6 is a generation-and-editing model, not only a text-to-image system.
• Multi-Image Reference Editing combines product, person, material, scene, and style references in one result.
• Web Grounding uses information from webpages and documents to add current context to an image prompt.
• The model supports Surgical Editing, including localized object changes, text updates, layout adjustments, and artifact cleanup.
• Arena reported 1,331 points and a No. 2 text-to-image ranking, with a reference cost of $38.90 per 1,000 output images.
• The MAI-Image-2.6-Flash variant reported a 12.2-second median response time in Microsoft's stated test, compared with 33.6 seconds for GPT-Image-2-Medium.

What Is MAI-Image-2.6?

MAI-Image-2.6 is an image foundation model developed by Microsoft AI. It handles both text-to-image generation and image-to-image editing, with a focus on photorealism, reliable text rendering, commercial imagery, portraits, 3D visuals, and cinematic scenes.

Microsoft announced the model on August 10, 2026, and later updated its launch material to state that MAI-Image-2.6 and MAI-Image-2.6-Flash were available in public preview on Microsoft Foundry. Microsoft also described the model as No. 2 on Arena's text-to-image leaderboard, with a 79-point improvement over MAI-Image-2.5 and a 91-point improvement in text rendering.

The model's public capability set extends beyond ordinary prompt-to-image generation. It can use several visual references, draw on webpage or document information, and make targeted changes to an existing image. It can also adapt layouts and aspect ratios for different output formats.

MAI-Image-2.6 is best understood as a generation-and-editing model, not only a text-to-image endpoint.

The model family includes a flagship version focused on quality and a Flash version focused on faster, more economical production. Results and pricing should therefore be labeled by variant rather than treated as interchangeable.

MAI-Image-2.6 at a Glance

Specification	Details
Developer	Microsoft AI
Type	Generative image model
Modality	Text-to-image and image-to-image editing
Input	Text and image
Output	Image
Context window	32,000 tokens listed for MAI-Image-2.6-Flash; flagship context not separately confirmed
Output resolution	Up to approximately 1.5K reported for the model family
Reference images	Up to 5 visual references reported in launch material
Pricing	About $38.90 per 1,000 output images in a reported Arena reference; exact universal rate not confirmed
Availability	Public Preview on Microsoft Foundry; MAI Playground and Arena access announced
License	Not yet confirmed
Public weights	Not yet confirmed
Exact API model ID	Not yet confirmed

The price in the table is a reference figure, not a promise that every endpoint or configuration charges the same amount. The available evidence also separates the flagship's benchmark price from Flash's reported lower-cost positioning.

How MAI-Image-2.6 Works / What Makes It Different

The strongest distinction is the combination of visual control and contextual input. A user can provide a written brief, several reference images, and supporting web information instead of encoding every visual constraint in prose.

Multi-Image Reference Editing

Multi-Image Reference Editing lets the model combine up to five visual references, according to the available launch material. Each image can contribute a different signal. One reference can define a product's shape, another its material, a third the environment, and a fourth the subject's pose.

A Microsoft example used several references for a sports shoe: a motion image, a canyon setting, a material close-up, a product shot, and a color palette. The intended workflow is compositional rather than purely imitative. The prompt tells the model which properties to retain and how to combine them.

This is useful for branded product scenes because a team can preserve recognizable product details while changing setting, lighting, and campaign direction.

Web Grounding

Web Grounding allows the model to use webpages and documents as context for an image request. Microsoft demonstrated the mode with Madrid's Metrópolis building, combining architectural details with a requested season, weather, lighting setup, and long-exposure effect.

Web Grounding does not mean the model continuously retrains on live information. The available material describes it as a generation-time context mechanism. Developers should still verify factual details, especially when a generated image represents a real building, person, location, or product.

Surgical Editing

Surgical Editing covers localized changes to an existing image. The documented operations include object removal, object replacement, attribute changes, local inpainting, text updates, enhancement, and motion-blur cleanup.

The goal is to change one region while preserving the rest of the composition. That makes the model relevant to product revisions, advertising variants, packaging changes, poster updates, and iterative art direction. Pixel-perfect preservation is not guaranteed, so production workflows still need visual review after each edit.

Dynamic Aspect Ratios and Rendering

Dynamic Aspect Ratios allow the model to adapt an image to different layouts rather than treating every request as a fixed square canvas. The reported output ceiling is approximately 1.5K resolution, although interfaces may expose different limits.

Microsoft also highlights stronger in-image text, portraits, 3D imagery, commercial design, and photorealistic scenes. These are capability claims supported by launch examples and benchmark positioning, not guarantees for every prompt language or typography style.

What You Can Do With MAI-Image-2.6

MAI-Image-2.6 is suited to workflows where generation and revision happen together:

• Product and branding imagery: Create commercial scenes from product references, color palettes, and material close-ups.
• Packaging and advertising: Generate layouts containing brand names, labels, product claims, and supporting design elements.
• Reference-driven art direction: Combine a subject, location, style, and lighting reference without describing each feature manually.
• Localized creative revisions: Replace an object, update poster text, remove blur, or alter a product attribute while retaining the broader composition.
• Real-world visualizations: Use Web Grounding for architectural, geographic, or documentary-style prompts that require external context.
• Batch concept development: Use MAI-Image-2.6-Flash for faster drafts, alternate compositions, and large-volume iteration.

The model is not limited to one visual style. Microsoft's materials cite portraits, 3D imagery, product design, cinematic imagery, and commercial scenes. Exact performance depends on the prompt, source images, language, and interface.

How MAI-Image-2.6 Compares

Arena reported MAI-Image-2.6 at No. 2 for text-to-image with 1,331 points and tied for No. 2 in image editing with 1,439 points. Artificial Analysis reported the model at No. 1 for image editing and No. 2 for text-to-image in its cited leaderboard updates. These rankings measure preference on particular evaluation sets, not every possible production task.

Model or variant	Reported result	Reported speed or price signal
MAI-Image-2.6	No. 2 text-to-image; tied No. 2 image editing	About $38.90 per 1,000 outputs; approximately $0.039 per image
GPT Image 2 high	1,318 image-editing Elo versus 1,324 for MAI-Image-2.6	About $211 per 1,000 outputs in the cited comparison
MAI-Image-2.6-Flash	1,311 image-editing Elo in the cited comparison	$19.50 per 1,000 outputs reported in the same reference chart; 12.2-second median latency in Microsoft's test

The comparison is not fully symmetrical. The quality-price chart uses GPT Image 2 high, while the latency comparison uses GPT-Image-2-Medium. Microsoft reported Flash as more than twice as fast in its selected test and claimed 72% lower GPU capacity use than GPT-Image-2. Those are vendor test claims, not an independent universal benchmark. Arena's published ranking is available in its text-to-image results.

For context on how other companies package image generation and editing into a model family, see the analysis of Meta's Muse Image and Muse Video.

At the reported Arena reference price, MAI-Image-2.6 costs about $0.039 per output image, but that figure should not be treated as a universal retail rate.

Availability: How to Access MAI-Image-2.6

The primary official access route is Microsoft Foundry, where MAI-Image-2.6 and MAI-Image-2.6-Flash are listed in public preview. Microsoft has also announced access through MAI Playground and Arena. The exact Foundry deployment settings, rate limits, billing tiers, and API model identifiers are not confirmed in the available materials.

Developers evaluating image pipelines can compare the workflow with GPT Image 2, which supports the same broad text-to-image and image-to-image task family. That link is a comparison option, not an indication that MAI-Image-2.6 is hosted on the same service.

Microsoft Foundry's model catalog lists Flash with text and image inputs, image output, a 32,000-token context window, and Preview lifecycle status. The catalog does not establish that every one of those technical fields applies identically to the flagship model.

What We Don't Know Yet

Several important engineering details remain open:

• The public materials do not confirm model parameter counts, training compute, or architecture details for the flagship.
• The exact API endpoint, model identifier, rate limits, batch behavior, and streaming behavior are not confirmed.
• The public weights and license status are not confirmed, so MAI-Image-2.6 should not be described as open source.
• Pricing differs between cited charts and variants. Flash is reported at lower cost, but the official universal price schedule is not established.
• Resolution reports vary by interface. Approximately 1.5K is the supported figure in the main capability discussion, while individual community reports mention different output sizes.
• Community reports describe inconsistent real-world wait times. Those reports are unconfirmed and do not replace a controlled latency test.

The most reliable current picture is therefore a hosted Microsoft model in public preview, with strong leaderboard placement and a broad editing feature set.

Frequently Asked Questions

What is MAI-Image-2.6?

MAI-Image-2.6 is Microsoft's generative image model for creating images from text and editing existing images. It supports multiple visual references, Web Grounding, localized edits, and outputs up to approximately 1.5K resolution.

Is MAI-Image-2.6 free?

MAI-Image-2.6 is available for public preview through Microsoft Foundry and can be tried through MAI Playground, but the bundle does not confirm a permanently free tier. Reported costs vary by model variant, interface, and usage.

How much does MAI-Image-2.6 cost?

A reported Arena reference price for MAI-Image-2.6 is about $38.90 per 1,000 output images, or approximately $0.039 per image. This is a benchmark price signal rather than a confirmed universal retail rate.

Is MAI-Image-2.6 open source?

MAI-Image-2.6 is not confirmed to be open source. The available materials describe Microsoft-hosted access through Foundry and related products, while a public weights license is not confirmed.

Where can I use MAI-Image-2.6?

MAI-Image-2.6 is available in public preview through Microsoft Foundry, and Microsoft says users can try it in MAI Playground and Arena. The exact API endpoint and model identifier are not confirmed in the available materials.

MAI-Image-2.6 vs GPT Image 2: which is better?

MAI-Image-2.6 is close to GPT Image 2 on the reported image-editing Elo comparison, trailing GPT Image 2 high by 6 points while costing less in the cited reference chart. Neither result establishes a universal winner across every prompt or workflow.

Does MAI-Image-2.6 support image editing?

MAI-Image-2.6 supports image editing, including object replacement, text updates, localized changes, layout adaptation, and artifact cleanup. It also supports combining multiple reference images in an editing or generation workflow.

What to watch next: track Microsoft's final Foundry pricing and API documentation, whether the model leaves public preview, and independent tests of Flash latency, output resolution, and multi-image consistency.

Building similar image-generation and editing workflows? On kie.ai you can try Nano Banana 2.1, Seedream 5.0 Pro, and Qwen Image 2.1.