Microsoft opens MAI Image 2.7 editing preview

Microsoft released MAI Image 2.7 in preview on October 27 for editing existing pictures. It takes an image and instructions to revise or restore existing assets. The company names Arena, Artificial Analysis and MAI Playground as access points; text-to-image generation is planned for later.

The public Playground linked from the announcement presents a more complicated entry point. On October 28, its address parameter and model selector still named MAI-Image-2.6, alongside a separate promotion for editing with 2.7. That establishes a promoted editing entry, while the selected generation model carries the earlier version number. It leaves access for a particular region or account unverified. This inspection involved reading the public page only: no account was used, picture uploaded or generation submitted. The page describes Playground as a limited preview of experimental models.

Microsoft describes changing an entire scene's lighting or style, moving products into another setting, adjusting text and symbols, and restoring assets. It reports gains over 2.6 across all ten evaluated use cases and seven editing actions, with a 24-Elo lead over the next model for scene and style.

Artificial Analysis's AA-Image-Editing v2.1 table, checked October 28, lists the preview second at 1,193 Elo. OpenAI's GPT Image 2.5 Sunburst scores 1,199 and Flare 1,182. The earlier MAI Image 2.6 scores 1,151, a 42-point difference. The preview's row records 4,891 samples.

Validated private evaluators make blind choices between edits produced from the same input image and instruction. The preview's 95% interval, 1,178–1,208, overlaps the leader's 1,188–1,210. These are preference rankings with visible uncertainty alongside the stated overall positions. The table also marks 2.7 as having no available API.

Arena's overall single-image editing board separately lists 2.7 third at 1,471, behind two OpenAI models. This explains Microsoft's distinction between third among models and second among laboratories. The preview has 3,402 votes, a displayed score margin of ±10, and a rank spread of second to fourth. Arena also provides pairwise win-rate and battle-count plots.

The board carries an October 23 date, separate from the October 28 inspection. Its roughly 31.57 million total votes cover 61 models; that aggregate is not the preview's own evaluation count. Confidence-interval plots use bootstrapping. These displayed counts and ranges accompany the model ranking, making the amount of evidence behind each row visible.

The preview and Foundry cover different versions

Microsoft separately made 2.6 and 2.6 Flash generally available in Foundry. The flagship accepts five reference images for consistent products, characters and brand assets. Family features include Bing context, dynamic aspect ratios and up to 1.5K output.

Foundry’s per-million-token starting prices for 2.6 are $5 for text input, $8 for image input and $38 for output; Flash costs $1.75, $2.50 and $19, respectively. These are token rates. Adobe confirms forthcoming 2.6 Firefly access; timing remains unspecified.