Microsoft Canvas Lets Designers Edit Images Region by Region for $10
Microsoft's new Canvas workspace runs on MAI Image 2.7 Flash. You can sketch, select and edit parts of an image in one place, at about a cent per image through the API.
- Microsoft launched Canvas, an MAI Labs workspace for sketching and selectively editing images.
- Canvas runs on MAI Image 2.7 Flash, launched with the flagship MAI Image 2.7.
- Pricing is about $10 per 1,000 images for Flash and $35 per 1,000 for the flagship in Foundry preview.
- MAI Image 2.7 ranks #1 on Artificial Analysis image editing, 4 Elo ahead of GPT Image 2.6 Sunburst.
- Both are 20B-parameter flow-matching diffusion models that support transparent backgrounds and output up to 2K.
- Available via Microsoft Foundry, fal, Genspark, OpenRouter and Vercel. Testers call the gain over 2.6 incremental.
Microsoft AI has launched Canvas, a browser-based image workspace in MAI Labs powered by MAI Image 2.7 Flash. Users can begin with a prompt or sketch, generate an image, and select regions to replace, move, or restyle without leaving the workspace. Canvas arrived alongside two proprietary models designed to make iterative editing faster and less expensive.
Canvas makes each region editable
Selecting an object or region confines an edit to that area, helping preserve the rest of the composition across successive revisions. The workflow supports tasks such as changing an object’s material, moving it within a scene, replacing a background, and turning a rough sketch into a finished diagram.
Transparent-background output also prepares product images, logos, icons, and illustrations for reuse without a separate cutout step.
Two models divide the workload
Microsoft positions MAI Image 2.7 for final-image quality and MAI Image 2.7 Flash for faster, lower-cost iteration. Flash powers Canvas because interactive sessions can involve many generations and localized edits.
| Model | Primary role | Preview price | Typical use |
|---|---|---|---|
| MAI Image 2.7 | Highest final quality | About $35 per 1,000 images | Final renders and quality-sensitive edits |
| MAI Image 2.7 Flash | Faster, cheaper generation | About $10 per 1,000 images | Drafts, variations, and high-volume workloads |
The model card provides the main technical specifications:
- Architecture: A diffusion model trained with flow matching for text-to-image generation and image-to-image editing. Flow matching teaches the model a direct path from random noise to a finished image, which can reduce the number of generation steps.
- Size: 20 billion non-embedding parameters across the MAI Image 2.7 family.
- Limits: A 32,000-token context window and output of about 4.2 million pixels, including 2,048 by 2,048 images.
- Availability: Microsoft does not publish the model weights. Usage is governed by the terms of the service hosting the model.
A leaderboard lead with little room
At launch, MAI Image 2.7 held first place on Artificial Analysis’s image-editing leaderboard. Microsoft reported an Elo score of 1,215 from 5,478 votes, four points ahead of GPT Image 2.6 Sunburst at its maximum setting. The company says its model costs one-fifth as much. A four-point margin can shift as the leaderboard collects more votes.
Microsoft also reports that MAI Image 2.7 improved on its predecessor across all 10 use cases and seven editing actions in the Artificial Analysis evaluation.
Microsoft’s launch materials quote Wirestock’s head of content saying the model “reads a scene the way a photographer does,” citing its handling of objects, lighting, angles, backgrounds, and fine details. Wirestock is a launch partner, so the comment represents partner feedback rather than an independent evaluation.
Pricing favors repeated drafts
Microsoft Foundry lists MAI Image 2.7 at about 3.5 cents per image and Flash at about one cent during public preview. Canvas remains in limited preview, and Microsoft has not published separate pricing for the workspace.
images_per_month = 50_000
flash_only = images_per_month / 1_000 * 10 # $500
flagship_only = images_per_month / 1_000 * 35 # $1,750
drafts = 45_000 / 1_000 * 10 # $450
finals = 5_000 / 1_000 * 35 # $175
mixed_total = drafts + finals # $625The mixed example routes 45,000 drafts through Flash and reserves the full model for 5,000 final renders. That pattern reduces generation costs for pipelines that produce many candidates before selecting an output.
APIs, model IDs, and access
Developers and designers can access the models through several services:
- Canvas: Interactive sketching, generation, and selective editing in MAI Labs.
- Microsoft Foundry: Preview APIs for generation and editing with both models.
- Partner platforms: fal, Genspark, OpenRouter, and Vercel.
Vercel AI Gateway exposes the models as microsoft/mai-image-2.7 and microsoft/mai-image-2.7-flash. Microsoft also identifies Adobe, Figma Weave, Minute, Wirestock, and Manus as partners. The broader MAI Image family already serves as the default model family behind Bing Image Creator.
Preview limits need planning
- No production SLA: Microsoft Foundry previews lack a service-level agreement, and Microsoft does not recommend them for production workloads.
- Fabricated details: The models can render readable text while inventing specifications, dates, labels, and component names. Technical diagrams require human verification.
- Quality and speed differ: The full model ranks higher for editing quality, while Flash targets lower latency, higher throughput, and lower cost.
- Closed deployment: Teams that require self-hosting, weight access, or model-level customization will need another option.
Route drafts and finals separately
Canvas fits sketch-to-diagram work, product mockups, background replacement, and localized material or color changes. API pipelines that generate many candidates can send drafts and variations to Flash, then use MAI Image 2.7 for selected final renders.
Evaluations on representative prompts should measure local-edit preservation, text accuracy, prompt adherence, latency, and total cost. Those results will determine whether the lower Flash price offsets any quality difference for a specific workload.