Artificial Analysis Revamps Image Editing Arena as Microsoft MAI Models Lead
Artificial Analysis announced on September 3 a major overhaul of its Image Editing Arena, expanding testing from single instructions to complex chained editing tasks across 17 fine-grained editing categories, ranking both the best overall model and the best model for each scenario. MAI-Image-2.6 tops the overall leaderboard.
Confirmed
- The evaluation is organized around 7 editing actions (e.g., object-level editing, identity-preserving editing, enhancement and restoration) and 10 real-world use cases (e.g., marketing ads, UI/UX design), for a total of 17 fine-grained categories.
- In the newly added "Enhancement & Restoration" category, Microsoft's MAI-Image-2.5, MAI-Image-2.6-Preview, and others swept the top four spots, while GPT Image 2 fell to sixth.
- In the newly added "Composition & Framing" category, GPT Image 2 leads Microsoft MAI by roughly 20 Elo; this category measures capabilities such as 2D/3D recomposition, changing camera angles, zooming out to expand a scene, and relayout while preserving design style.
- Across the fine-grained evaluations, MAI leads in 7 categories and GPT takes 5, with each model excelling in different scenarios.
- The arena has opened human voting and published its full methodology, including prompt writing, refresh and elimination mechanisms, and how Elo scores are computed; the price-quality frontier data shows image editing service prices ranging from $20 to $211 per thousand images.
Why it matters
- The revamped arena shifts from "which is the best model" to "which model for which scenario," providing more granular guidance for use-case-based model selection.
- Publishing the price-quality frontier makes cost-versus-quality tradeoffs quantifiable, offering direct guidance for commercial deployment choices.
- The Microsoft MAI series' strong showing on the overall leaderboard and in multiple subcategories shows it is posing substantive competitive pressure on OpenAI's GPT Image series in image editing.
2026-09-03 ~ 2026-09-03 · 5 related posts
Primary sources
- Image Editing Arena revamp: MAI-Image-2.6 leads 7 of 17 categories, GPT Image 2 takes 5 — ArtificialAnlys ·
- GPT Image 2 leads MAI-Image-2.6 by 20 Elo on new Composition & Framing editing benchmark — ArtificialAnlys ·
- Microsoft MAI image models sweep top 4 on Enhancement & Restoration; GPT Image 2 falls to #6 — ArtificialAnlys ·
- Image Editing Arena Revamp: MAI-Image-2.6 Leads, GPT Image 2 Best at Local Edits — ArtificialAnlys · 2026-09-03
- [source] Image Editing Arena revamp: MAI-Image-2.6 leads 7 of 17 categories, GPT Image 2 takes 5 — ArtificialAnlys · 2026-09-03
- [source] Microsoft MAI image models sweep top 4 on Enhancement & Restoration; GPT Image 2 falls to #6 — ArtificialAnlys · 2026-09-03
- [source] GPT Image 2 leads MAI-Image-2.6 by 20 Elo on new Composition & Framing editing benchmark — ArtificialAnlys · 2026-09-03
- Image Editing Arena opens with a $20-to-$211 price-quality frontier — ArtificialAnlys · 2026-09-03