Microsoft's MAI-Image-2.6 lands at #2, 45 Elo behind GPT Image 2

Microsoft's in-house image model just took #2 on Arena's text-to-image leaderboard with 1336 Elo — 45 points behind GPT Image 2 (Medium) at the top, and 20 ahead of Grok Imagine Image 2.0 at #3. It also clears Google's Nano Banana and Meta's models.

The more telling number is the jump from its own predecessor: +79 Elo over MAI-Image-2.5, with gains in all 8 Arena categories. The biggest movers were 3D (+114), portraits (+105) and text rendering (+91) — exactly the three things image models are usually worst at, and exactly what commercial work needs. Microsoft is pitching it at product shots, branding and cinematic output, plus multi-reference inputs and more control over format and resolution.

You can try it on Arena now. MAI Playground gets it later this week, with Microsoft Foundry and API access rolling out after — Mustafa Suleyman's team going from third place to second in one release cycle.

Why it matters: Microsoft is quietly building a frontier image stack it doesn't have to license from OpenAI.