MODEL PROFILE · Microsoft AI

MAI-Image-2.6

Microsoft’s MAI-Image-2.6 preview supports reference editing and web grounding. Review Foundry limits, route-specific pricing and evaluation tradeoffs.

Source reviewed: 2026-10-04 · Image generation and editing · public preview

Provider documentation summary. No hands-on results or independent ranking are claimed.

Documented facts

The following fields come from the official provider source, reviewed on the date above.

Exact version: MAI-Image-2.6, version 2026-07-31, Global Standard deployment in Microsoft Foundry.
Inputs and output: Text-to-image and image editing; up to five JPEG/PNG references, PNG output.
Image size: At least 768 pixels per dimension; maximum 2,359,296 total pixels, equivalent to 1536×1536. Rectangles may exceed 1536 on one side within that area limit.
Prompt limit: The Foundry guide lists a maximum context length of 32,000 tokens.
Controls: auto_aspect_ratio selects framing; web_grounding can add Bing Search context.
Preview terms: No service-level agreement; Microsoft does not recommend preview for production workloads.

Dated access and model positioning

Microsoft announced Foundry public-preview access on September 4, 2026, alongside MAI-Image-2.6-Flash. Its launch post offers both in MAI Playground and positions the standard model for precision and Flash for faster production. Those are vendor descriptions, not a DualView speed or quality measurement. The announcement’s leaderboard positions are explicitly dated September 4; they should not be treated as current rankings. Microsoft’s September 4 announcement.

Foundry API identity

Use /mai/v1/images/generations or /mai/v1/images/edits on your Azure resource. The model field is your deployment name, not necessarily the catalog name. Editing uses multipart form data. Access requires a paid Azure subscription, deployment permissions and a supported region. Microsoft deployment and request guide.

Listed pricing: OpenRouter route

OpenRouter’s MAI-Image-2.6 page lists $5 per million text-input tokens, $8 per million image-input tokens and $38 per million image-output tokens. For a hypothetical request billed for 1,000 text tokens and 1,000 output-image tokens, with no image inputs or other fees, the arithmetic is $0.005 + $0.038 = $0.043. This is a token-budget example, not a fixed price per image or a direct-Azure quote. Its page displays 4K context, which differs from the Foundry guide; verify the selected route’s actual limits before migration. OpenRouter model listing and rate table.

The practical decision

Evaluate this model when an image workflow needs both composition control and reference-based revisions. Treat the choice between standard and Flash as a question about accepted output quality at your own budget. A campaign may need exact lettering or a product’s shape to survive an edit; another may only need a rough layout. Write those acceptance conditions before reviewing candidates so a visually attractive result cannot conceal a failed requirement.

What to record in your own comparison

Keep the source references and edit request identical across candidates. Inspect the requested change separately from details that should remain untouched, especially labels, logos, edges and lighting. Record both kinds of failure.
Compare web grounding enabled and disabled as separate conditions. Check factual details against the intended source; a plausible image is not verification of a current event, location or product. Preserve the review date and any evidence used.
For fixed-layout work, choose one framing requirement and assess whether useful content was cropped, stretched or rearranged. Compare files at native size before resizing them into a shared presentation canvas.
Record the exact model, route, deployment version, dimensions and billed input/output token counts. Include failed attempts in the cost per accepted asset. A cheaper successful call may still produce a more expensive deliverable if it requires more retries.
Use another image-generation/editing model for the same task, such as FLUX.3 Image, and keep any difference in reference count or available controls visible. Avoid turning one successful example into an overall ranking.

Limitations and unknowns

This is a documentation-based profile, not a hands-on test. We did not call either service or measure fidelity, grounding accuracy, latency or reliability. An OpenRouter listing does not prove access for a particular account; its public models API did not return this identifier during this review. Direct Foundry pricing, downloadable weights and a weight license were not established from the cited sources. The version date is distinct from the public-preview announcement and this review date. Preview access is not general availability.

Compare your saved results

Use the DualView editor to inspect files you already have. The accepted-output calculator uses your own rates and counts. Neither link runs a model or purchases generation.

Related reporting

Related model profiles

All model profiles and dated changes · Subscribe via RSS