Evaluating Microsoft MAI-Image-2.5: Photorealism and Spatial Editing

TL;DR: Microsoft's June 2026 release of MAI-Image-2.5 delivers a 75-point Elo gain over MAI-Image-2, securing the No. 2 spot on the Arena Image Edit leaderboard. Our hands-on testing of its complex spatial reasoning and text rendering demonstrates that this model offers enterprise-grade image generation and editing at a highly competitive price point of $47 per million output tokens.

Microsoft launched MAI-Image-2.5 on June 1, 2026, establishing a new benchmark for enterprise visual generation by securing the No. 2 rank on the Arena Image Edit leaderboard. This release addresses the limitations of previous models by introducing advanced spatial reasoning and precise localized editing features. For a broader look at the competitive environment, See our Full Guide. Our technical evaluation of the model confirms its ability to handle complex prompts that require strict adherence to spatial coordinates, light sources, and embedded text.

How does MAI-Image-2.5 compare to competitor models on Arena benchmarks?

MAI-Image-2.5 outperforms GPT-Image-1.5 and Nano Banana Pro 2K on standard Arena benchmarks, ranking No. 3 globally for text-to-image generation and No. 2 for image editing. The model registers a 75-point Elo improvement over its predecessor, MAI-Image-2. The largest performance gains appear in Text Rendering, which saw a 107-point increase, followed by Cartoon, Anime & Fantasy with a 90-point gain.

Analyzing the Win Rates Against Nano Banana 2.1

In blind human preference testing conducted between May 31 and June 1, 2026, MAI-Image-2.5 achieved a net advantage across 12 distinct editing categories. It consistently outperformed Nano Banana 2.1, particularly in complex evaluations requiring more than 100 judged matches. The evaluation metrics excluded low-quality outputs from both models, focusing strictly on high-fidelity generations where MAI-Image-2.5 demonstrated superior prompt adherence and visual coherence.

Advanced spatial reasoning enables precise localized image editing

MAI-Image-2.5 understands scene structure, lighting, scale, and spatial relationships to execute highly targeted localized edits without altering the rest of the image. Enterprise users often need to modify specific elements of an asset rather than regenerating the entire image. MAI-Image-2.5 addresses this requirement by supporting precise edits such as replacing an object, correcting text, or removing motion blur. The model preserves the original scene's context, calculating perspective and casting accurate shadows for newly added elements.

Facial identity preservation across edits

A common failure point in automated image editing is the degradation of human facial features during scene modifications. MAI-Image-2.5 maintains recognizable likeness and facial identity across changes in pose, expression, or viewpoint. This capability allows marketing teams to reuse character assets consistently across different campaign materials without manual retouching.

What is the pricing and integration roadmap for MAI-Image-2.5 in enterprise workflows?

Microsoft has integrated MAI-Image-2.5 directly into PowerPoint and OneDrive, while offering developers API access through Foundry starting at $5 per million text input tokens. PowerPoint users can generate presentation-ready visuals directly from text prompts to accelerate slide creation. OneDrive users receive precise photo-editing capabilities, including background cleanup and distraction removal. For custom deployments, Microsoft offers two API tiers: the premium MAI-Image-2.5 and the high-speed MAI-Image-2.5-Flash.

Token pricing breakdown for developers

The premium MAI-Image-2.5 model costs $5 per 1 million text input tokens, $8 per 1 million image input tokens, and $47 per 1 million image output tokens. For high-volume production workloads, MAI-Image-2.5-Flash reduces costs to $1.75 per 1 million text inputs, $1.75 per 1 million image inputs, and $19.50 per 1 million image outputs. These pricing structures allow engineering teams to balance cost, processing speed, and visual fidelity depending on their specific application needs.

How does the model mitigate content generation risks?

MAI-Image-2.5 incorporates layered safety guardrails, including prompt and output filtering, to actively detect and block policy-violating content. While these filters mitigate risk, Microsoft advises that the model can still reflect biases present in its training data or produce plausible but inaccurate visual details. Consequently, organizations must implement human review processes before deploying these generated images in high-stakes or sensitive contexts. This requirement applies specifically to legal, medical, financial, or news-related publishing workflows where visual accuracy and identity verification are critical.

Key Takeaways

  • MAI-Image-2.5 ranks No. 2 on the Arena Image Edit leaderboard, delivering a 75-point overall Elo improvement over MAI-Image-2 as of June 2026.
  • The model introduces advanced spatial awareness, enabling localized edits like object replacement and motion blur removal while preserving original scene lighting and perspective.
  • Developers can access the models via Foundry, with the premium tier priced at $47 per million image output tokens and the Flash tier at $19.50 per million image output tokens.