Microsoft's MAI-Image-2.5-Pro Just Topped an AI Image Editing Leaderboard

Microsoft's MAI-Image-2.5-Pro debuted at number one on Artificial Analysis's Image Editing Leaderboard, surpassing Google, OpenAI's GPT Image 2, and Microsoft's own MAI-Image-2.5. The model, launched in preview on Microsoft Foundry on July 23, also ranks seventh in Text to Image. Pricing is set at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens, roughly $108.50 per thousand 1024x1024 images.

Microsoft's newest image model debuted at number one on Artificial Analysis's image editing leaderboard this week, beating Google, OpenAI and its own prior model in the process. MAI-Image-2.5-Pro launched in preview on Microsoft Foundry on July 23, and less than a month later it's sitting at the top of a leaderboard that matters. According to a post from Artificial Analysis on X, the model debuts at number one on the firm's Image Editing Leaderboard and lands at number seven in Text to Image. That's a fast climb for a model most builders hadn't heard of a month ago. The editing win is the headline here, and it's a real one. Artificial Analysis runs its Image Arena by pitting models head to head and letting outcomes settle into an Elo-style ranking, the same approach LMArena popularized for chatbots. MAI-Image-2.5-Pro's editing score surpasses Reve 2.1, GPT Image 2, and Microsoft's own MAI-Image-2.5, which had itself launched at number two on the same board only weeks earlier. Beating your predecessor is one thing. Beating OpenAI's flagship image editor is another. Pro is the top of a three-model family that now includes MAI-Image-2.5 and MAI-Image-2.5-Flash. Microsoft frames the trio as covering the quality-speed-cost curve, so a developer building a product can pick the model that fits the job rather than paying premium rates for a quick thumbnail. Flash is for volume and speed. The standard 2.5 model is the balanced pick. Pro is different. Microsoft calls it its highest-fidelity image model to date, built for hero imagery, detailed editing work, and precise text rendering inside a generated image, the kind of thing that trips up most image models badly. Text rendering is a small detail with a big payoff. Ask most image generators to put a specific sentence on a sign or a product label and you'll get gibberish half the time. Microsoft is explicitly positioning Pro to fix that, which suggests the target customer isn't a hobbyist making art but a company that needs clean, on-brand creative at scale, ad banners, packaging mockups, marketing assets with real copy on them. DeepSeek Warns Its Ultra-Cheap V4-Flash Pricing Won't Last https://startupfortune.com/deepseek-warns-its-ultra-cheap-v4-flash-pricing-wont-last/ DeepSeek's new V4-Flash model is the cheapest frontier-grade AI ever benchmarked by Artificial Analysis, costing about three cents to run a full test suite. Days after launch, DeepSeek warned developers of a coming 'significant' price hike without naming a date. - deepseek v4 flash pricing https://startupfortune.com/deepseek-warns-its-ultra-cheap-v4-flash-pricing-wont-last/ - cheapest ai api models 2026 https://startupfortune.com/deepseek-warns-its-ultra-cheap-v4-flash-pricing-wont-last/ The pricing tells the same story. On Microsoft Foundry, MAI-Image-2.5-Pro runs $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens. Translated into something a marketing team would actually budget against, that works out to roughly $108.50 per thousand 1024x1024 images. Compare that to $48 per thousand images for the standard MAI-Image-2.5 and $20 per thousand for Flash. The tiering is obvious: you pay more than double the mid-tier price and more than five times the budget tier to get the model that just beat GPT Image 2 on editing quality. Frankly, that's a steep premium. Microsoft is betting builders will pay it for the jobs where quality actually matters, rather than run everything through Pro by default. Where this fits in Microsoft's AI push The MAI brand is Microsoft's own in-house model effort, separate from the OpenAI models it also resells through Azure. That's not a footnote. Building a family that now ranks at the top of an independent image benchmark is a signal Microsoft wants to be taken seriously as a model developer in its own right, not just an OpenAI distributor with a cloud around it. Landing number one in editing while OpenAI's GPT Image 2 sits below it on the same board is exactly the kind of result that gets attention inside enterprise AI budgets, where editing tools for product photography and ad creative are a bigger recurring cost than one-off text-to-image generation. That's also probably why the number seven finish in Text to Image doesn't dent the story much. Pro wasn't built to win pure generation from a blank prompt. It was built to take an existing image and make precise, high-fidelity changes to it, and that's the leaderboard it topped. MAI-Image-2.5-Pro is currently available in preview across seven global-standard regions on Microsoft Foundry, and anyone can try it directly through the MAI Playground without writing a line of code. For a model three weeks old, that's a fast path from launch to leaderboard to public hands. Also read: Reddit Users Say OpenAI Quietly Cut Codex Usage Limits in Half https://startupfortune.com/reddit-users-say-openai-quietly-cut-codex-usage-limits-in-half/ • AI Agents Autonomously Hacked Taiwan's Government for Four Days Straight https://startupfortune.com/ai-agents-autonomously-hacked-taiwans-government-for-four-days-straight/ • Moonshot AI's Kimi K3 Undercuts GPT and Claude While Proving Its Numbers https://startupfortune.com/moonshot-ais-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers/