Microsoft's newest image model ranks second on a major AI leaderboard

Microsoft said its latest image-generation model, MAI-Image-2.6, launched in second place on the Arena text-to-image leaderboard — the crowdsourced ranking run by the platform formerly known as LMArena — ahead of entries from Google, Meta and xAI. Microsoft did not name which model held first place, and the leaderboard position reflects Arena’s own methodology of anonymous head-to-head user votes rather than an independent or standardised benchmark.

The company reported gains specifically in text rendering within images, an area where diffusion models have historically struggled, alongside improvements it described in portraiture, 3D imagery and photorealistic commercial output. Against its own predecessor, MAI-Image-2.6, Microsoft said the model gained about 79 Elo points overall and about 91 Elo points on text-rendering tasks specifically — self-reported figures, since Elo scores on Arena are a function of the live leaderboard rather than a fixed test set. The model was made available on Arena itself first, with a rollout to Microsoft’s own Foundry platform and other products to follow.

The release is notable less for the ranking itself, which shifts as new entrants are added, than for what it signals about Microsoft’s product strategy: alongside its in-house MAI-Thinking-1 reasoning model, it is building a broader stable of models developed independently of OpenAI, the partner whose systems have otherwise powered most of Microsoft’s AI products since 2023.