MAI-Image-2.5 Guide: Production Status, Pro Preview, and Evaluation

MAI-Image-2.5 is Microsoft’s production image generation and editing model. In its July 23, 2026 update, Microsoft said the model had become the default engine powering Bing Image Creator end to end and was also running image workflows in PowerPoint and OneDrive. The same update placed the separate MAI-Image-2.5-Pro variant in public preview. Those are Microsoft source claims, not results produced by MAI Image Pro.

Disclosure: This article is an independent, source-backed guide. It is not a hands-on benchmark, and MAI Image Pro is not affiliated with Microsoft. The small images on this site were generated with a local Z-Image Turbo workflow and are illustrative assets, not MAI-Image-2.5 outputs. See our editorial and testing policy.

Key takeaways

• Microsoft reports that MAI-Image-2.5 is in production and now powers Bing Image Creator end to end as its default model.
• Microsoft also reports production use in PowerPoint image-to-image workflows and key OneDrive image-editing scenarios.
• MAI-Image-2.5-Pro entered public preview on July 23, 2026 as Microsoft’s highest-fidelity image variant; it is not the same model label as MAI-Image-2.5.
• The earlier June Arena announcement remains useful historical benchmark context, but it should not be used to imply that the model is still only announced or unreleased.
• A useful independent evaluation must publish prompts, settings, sample counts, failures, dates, and unedited outputs. This page does not claim those tests have been run.
• Rankings are time-sensitive. Verify the live text-to-image leaderboard and image-edit leaderboard before quoting a current position.

What is MAI-Image-2.5?

MAI-Image-2.5 is a Microsoft model for creating and editing images from text or photo prompts. The official model page emphasizes photorealistic output, design-ready imagery, precise editing, natural lighting, accurate skin tones, and closer control over requested changes. Microsoft’s July production update documents current product use, while the earlier Arena announcement provides dated benchmark context.

The name needs careful attribution. MAI-Image-2.5 is the Microsoft model. MAI Image Pro is this independent guide and browser workspace. It should not be described as a Microsoft product, official Microsoft property, or source of Microsoft’s benchmark data.

Availability needs precise language, not just a date. The June article documented an earlier Arena milestone. By July 23, Microsoft described MAI-Image-2.5 as an existing production model and MAI-Image-2.5-Pro as a new public-preview variant. We corrected this guide on July 28, 2026 so “announced,” “production,” and “public preview” are not treated as interchangeable statuses.

What did Microsoft’s earlier Arena announcement report?

In the earlier June article, Microsoft reported two different Arena positions: No. 3 for text-to-image and No. 2 for image editing. These rankings refer to separate, dated leaderboards and should not be collapsed into a generic current claim that the model is simply “No. 2” or “No. 3.” The announcement also says MAI-Image-2.5 exceeded MAI-Image-2 by 75 overall Arena points in the comparison shown as of June 1, 2026.

Public claim	Reported result	Scope and date
Text-to-image position	No. 3	Microsoft announcement; evaluation current around June 1, 2026
Image-editing position	No. 2	Microsoft announcement; Arena image-edit leaderboard
Improvement over MAI-Image-2	+75 Arena points	Microsoft’s comparison figure, not an MAI Image Pro test
Largest highlighted gain	Text rendering	Microsoft’s category comparison

Arena uses blind human preference comparisons, which makes it useful for measuring perceived output quality. It does not prove that one model is best for every brand, prompt language, aspect ratio, typography task, or production constraint. A leaderboard position can also move when new models enter, votes accumulate, or evaluation methods change.

For that reason, cite the dated Microsoft statement when discussing the launch result and cite the live Arena page when discussing the current ranking. The two statements answer different questions.

What changed from MAI-Image-2 to MAI-Image-2.5?

Microsoft’s public comparison emphasizes a broad quality improvement rather than one isolated feature. The most concrete reported difference is the 75-point overall Arena improvement, with text rendering identified as the largest category gain. Microsoft also highlights stronger photorealism, instruction following, editing coherence, and design-oriented output.

The following comparison separates sourced claims from practical evaluation questions:

Area	Microsoft’s public description	What an independent test should measure
Text rendering	Largest reported category gain	Character accuracy, misspellings, layout consistency, and legibility at final size
Prompt following	Closer control over the requested image	Required-object recall, forbidden-object avoidance, composition, and camera direction
Image editing	Precise edits that stay coherent in context	Identity retention, background preservation, local edit boundaries, and unintended changes
Photorealism	Natural lighting and accurate skin tones	Skin texture, hand anatomy, reflections, material response, and lighting continuity
Design-ready imagery	Stronger commercial and layout-oriented output	Logo-safe space, product geometry, label fidelity, crop flexibility, and downstream edit effort

This distinction matters because a product claim is not the same as a reproducible result. Until a test publishes its complete inputs and unedited outputs, it should be described as an evaluation plan rather than evidence.

How should MAI-Image-2.5 be evaluated fairly?

A fair evaluation uses a fixed prompt set, records every output, and declares the scoring rules before generation. Cherry-picking one attractive image answers whether the model can produce a good result once; it does not answer how reliably the model follows instructions across repeated attempts.

Use this minimum protocol:

1. Record the environment. Save the access route, visible model version, test date, region, aspect ratio, quality mode, edit strength, seed controls, and credit or price information.
2. Freeze the prompt set. Use the same prompt wording across all compared models. If a provider requires model-specific syntax, publish both the original prompt and the adapted version.
3. Generate repeated samples. Use at least four outputs per prompt when the interface allows it. Report all outputs, not only the strongest one.
4. Separate task categories. Score text-to-image, image editing, typography, product imagery, portraits, spatial instructions, and style control independently.
5. Define failure criteria. Mark missing objects, unreadable required text, incorrect counts, unintended edits, broken anatomy, and unusable crop constraints.
6. Keep source files. Publish unedited outputs or checksums, along with prompts and settings, so another reviewer can inspect the same evidence.
7. Report uncertainty. Distinguish consistent behavior from a result observed in one prompt or one generation batch.

MAI Image Pro has not completed that protocol for MAI-Image-2.5 at the time of this update. We are publishing the method now so future tests can be assessed against a declared standard instead of changing the rules after seeing outputs.

Which prompt structure is useful for evaluation?

A useful prompt separates required content from visual direction and constraints. This makes failures easier to diagnose than a long paragraph where subject, style, camera, text, and exclusions are mixed together.

Use the following structure:

Task: [text-to-image or edit]
Primary subject: [person, product, place, or object]
Required elements: [objects, count, pose, exact text]
Composition: [camera angle, crop, hierarchy, negative space]
Lighting and material: [source, softness, reflections, texture]
Style: [photographic, editorial, illustration, packaging, diagram]
Output use: [poster, product page, slide, social ad]
Constraints: [what must remain unchanged or must not appear]

For typography, put the exact required text on its own line and preserve punctuation and capitalization. Then score the result character by character. “Looks readable” is too vague for a text-rendering comparison; an error count is more useful.

For image editing, identify both the change and the protected regions. A prompt such as “replace the red label with a blue label; preserve bottle geometry, reflections, background, camera angle, and all other text” gives the evaluator explicit boundaries.

For product imagery, include the output channel and crop. A square marketplace image, a 9:16 social asset, and a wide landing-page hero impose different composition requirements even when the subject is identical.

Where can MAI-Image-2.5 be accessed?

Microsoft links to its AI Playground from the official model page, and the model also has a Microsoft Foundry catalog entry. The July 23 update says MAI-Image-2.5-Pro is in public preview and links to Foundry, Microsoft Learn, and MAI Playground. Availability, regions, quotas, pricing, preview status, and account requirements can change, so the current Microsoft interfaces are the authority for access conditions.

MAI Image Pro also provides a browser workspace and published pricing information. That workspace is operated independently. Its plans, credits, and interface should not be confused with Microsoft Foundry pricing or official Microsoft access terms.

Before committing a production workflow, verify:

• whether the interface identifies the exact model version;
• whether image editing and text-to-image are both available;
• the current resolution, aspect-ratio, upload, and export limits;
• regional or account restrictions;
• the commercial-use terms that apply to the chosen provider;
• how prompts, uploads, and generated assets are retained.

What are the main limitations of this guide?

This guide does not establish real-world quality, latency, cost, or reliability for MAI-Image-2.5 because it does not publish a completed hands-on test dataset. It summarizes public primary sources and provides an evaluation method. Readers should not infer unpublished test results from the presence of illustrative images on the site.

The article also does not treat Arena as a universal product score. Arena results capture preference under the platform’s evaluation conditions. Production teams still need task-specific checks for exact text, brand constraints, repeated characters, identity preservation, licensing, latency, and cost.

Finally, model pages and rankings are time-sensitive. This article records what the cited sources said when reviewed on July 28, 2026. Send a correction to [email protected] if a source moves, a ranking changes, or a sentence no longer matches the primary evidence.

Primary sources and further reading

1. Microsoft AI, “Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash” — published July 23, 2026; production use and Pro public-preview status.
2. Microsoft AI model page for MAI-Image-2.5 — model family, capabilities, model cards, and official Playground link.
3. Microsoft AI, “MAI-Image-2.5 launches at No. 2 for image editing on Arena” — earlier dated benchmark context.
4. Microsoft Foundry catalog entry for MAI-Image-2.5 — current catalog availability.
5. Arena text-to-image leaderboard — live ranking reference.
6. Arena image-edit leaderboard — live image-edit ranking reference.
7. About MAI Image Pro — publisher identity and independence disclosure.
8. MAI Image Pro editorial policy — sourcing, testing, media labeling, and corrections.