Skip to content

Models & products

MAI Image 2.7 adds generation and Flash at about $10 per 1,000 image outputs

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. MAI Image 2.7 and Flash reach Foundry public preview and partner APIs.
  2. Fal quotes $0.01024 for a 1024×1024 Flash output, with inputs billed separately.
  3. Flash powers the new Canvas workspace; MAI Labs is for personal, non-commercial experimentation.
PracticalDesignersDevelopers
Conceptual worktable with a botanical sketch, floral print and blank tablet
AI-generated conceptual image: a creative desk connecting sketching with image creation and revision · AI Ping

Microsoft introduced MAI Image 2.7 and the lower-cost MAI Image 2.7 Flash on November 17. Both generate and edit images and are in public preview on Microsoft Foundry, with partner distribution through fal, Genspark, OpenRouter and Vercel. Microsoft quotes approximately $35 per 1,000 images for the full model and $10 for Flash.

The October 27 preview started with editing existing images, while text-to-image generation was promised later. The new release adds generation and a Flash option. Microsoft positions the full model for precise edits and Flash for speed and cost efficiency.

Editing requests can change an object’s color, material or position while retaining the surrounding composition. The model card lists removal, replacement, inpainting, text changes and artifact cleanup. Microsoft says the model is designed to preserve visual consistency through successive edits.

Integration limits differ. The MAI Labs model interface accepts up to five reference images, while fal’s editing endpoint specifies up to two inputs. The model card caps output at 4,194,304 pixels, equivalent to 2048×2048. Either dimension may exceed 2048 if the total stays within that cap.

Conceptual hand holding a coral swatch beside a blue lamp and material samples
AI-generated conceptual image: material swatches represent targeted color and texture edits · AI Ping

Both models support transparent backgrounds. Microsoft describes uses including isolated product images, logos, icons and illustrations. Fal lists PNG, JPEG and WebP as its available output formats.

Conceptual teapot cutout on a clear panel and a plum folder
AI-generated conceptual image: clear acrylic and a cutout represent separating a product from its background · AI Ping

Input costs sit outside the 1,000-image headline

On fal, the full model costs $5 per million text-input tokens, $8 per million image-input tokens and $34 per million image-output tokens. A 1024×1024 output uses 1,024 tokens and costs $0.034816 before inputs. Its example with 1,024 image tokens on each side costs $0.043008 for image tokens, plus the text input.

Flash charges $1.75, $2.50 and $10 per million tokens for those respective categories. A 1024×1024 output costs $0.01024 before inputs. Resolution and reference images therefore affect the bill behind the roughly $10-per-1,000 figure. Fal rounds each total request cost up to the nearest $0.0001.

Artificial Analysis’s editing board, checked November 17, places the full model first at 1,215 Elo across 5,478 comparisons. Its 95% interval, 1,201–1,229, overlaps the second-ranked GPT Image 2.6 Sunburst maximum setting’s 1,198–1,224 interval. Evaluators compare outputs from the same source image and instruction without seeing the model names.

Text-to-image results are separate. The full model ranks fourth at 1,214 Elo over 5,008 comparisons; Flash ranks eleventh at 1,158 over 5,951. The full model’s displayed rank range is third to fifth. These are preference comparisons of newly generated images from matched prompts, a different task from editing an existing image.

Microsoft also launched Canvas, a Flash-powered workspace in MAI Labs. Users can begin with a prompt or sketch and select part of an image to replace, move or restyle, continuing revisions in one workspace.

Labs is reserved for personal, non-commercial experimentation. Its FAQ excludes work and school use and says availability varies by country. A Microsoft account is required, and the model interface states an age threshold of 18. The FAQ says uploaded files and generated content are retained for up to 29 days. Prompts and outputs may support service operations and improvements; Microsoft says it currently excludes them from model training.

Sources