AI DevelopmentNew Release8 min readPublished

Microsoft MAI-Image-2.7 and Flash: Editing, Cost, Access

At a 1,024-pixel square, the precision model adds roughly 2.5 cents per output. The harder buying decision is which edits need it and where the resulting images can be used.

DA
Digital Applied
PublishedNovember 19, 2026
Read time8 min

Changing the background behind a product is an easy brief to write. Keeping every letter on its label and every reflection on its glass intact is the expensive part. Microsoft released MAI-Image-2.7 and MAI-Image-2.7 Flash in public preview on November 17. It quotes about $35 per 1,000 images for the precision model and $10 for Flash. For a team making catalogue variations, the extra 2.5 cents a picture can be worth paying whenever it saves even a little skilled retouching.

The service you open also determines what you can do with the result. Canvas in MAI Labs runs Flash and is for personal experimentation. A designer preparing a paid brief needs a business route even for the first mock-up. Foundry offers an Azure preview for evaluation. On the partner side, fal explicitly marks its two editing endpoints for commercial use. The model name alone answers neither the pricing nor the permissions question. Microsoft’s November 17 announcement

Key takeaways
  1. 01
    Flash is the first candidate for disposable variationsIt supports generation and editing at a lower unit price. Start the precision model on work where preserving exact product features or typography is a condition of acceptance.
  2. 02
    The independent price comparison is for a 1,024 × 1,024 outputArtificial Analysis lists $34.82 and $10.24 per 1,000 default-setting square generations. Other resolutions, editing inputs and partner billing can change the bill.
  3. 03
    The first-place editing result belongs to the precision modelOn November 19, Artificial Analysis listed MAI-Image-2.7 first by Elo and Flash seventh. The top two models’ score intervals overlapped.
  4. 04
    Choose the work surface before supplying a client referenceMAI Labs restricts use to personal experiments. Azure has preview-specific terms and partners set their own conditions for the endpoints they sell.

01 — Creative briefWhat needs to stay put?

Microsoft describes MAI-Image-2.7 as the precision model for localized changes, text updates and consistent edits across successive turns. Flash is its optimized, lower-cost variant. Both can start with a text prompt or change an existing image. Both support transparent backgrounds. The difference worth testing is how often each preserves the material you never asked it to change. Microsoft advertises faster generation for Flash. Its launch announcement gives no latency figure that would tell a designer how long a particular edit will take.

Consider two requests for the same shoe. In the first, a designer wants twelve backgrounds for a mood board, with freedom to vary the lighting. In the second, the ecommerce team supplies a photographed shoe and requires the eyelets, seam pattern, logo placement and sole profile to match the physical item. Flash is a sensible first choice for the mood board. On the product image, run the precision model against the original photograph and ask a reviewer to inspect the unchanged details. A convincing synthetic shoe can still be the wrong shoe.

The joint model card describes prompt enhancement before the image model receives the input. If a concept repeatedly drifts, give the editor a narrow change and an explicit list of features to preserve. Save the original and each accepted intermediate. When a third edit distorts the logo, the second accepted version is a much better starting point than a fourth request to repair the third. Microsoft’s joint model card

02 — Price unitsWhat does $35 cover?

Microsoft gives rounded launch estimates for Foundry. Artificial Analysis provides a more tightly defined comparison on its editing leaderboard. The price column there is the cost of generating 1,000 images at 1,024 × 1,024 on the creator’s API using default settings. It sits alongside editing scores, but it does not price the reference material or a sequence of revisions for a particular edit job.

US dollars as of November 19, 2026. The first two rows use 1,000 image outputs. The final two rows use one million tokens. A provider’s input and output meters are separate.
Source and unitMAI-Image-2.72.7 Flash
Microsoft launch estimate / 1,000 images~$35~$10
Artificial Analysis / 1,000 square generations$34.82$10.24
OpenRouter / 1M prompt tokens$5.00$1.75
OpenRouter / 1M image-output tokens$34.00$10.00

On OpenRouter, the image-output number is a token rate. Multiplying $34 by the number of finished pictures gives the wrong units. Use billed image-output tokens divided by one million, multiplied by $34, then add the applicable input charges. An editing request can also carry reference images. Keep the provider’s usage breakdown for a representative request before estimating an entire catalogue. OpenRouter precision model · OpenRouter Flash model

Azure’s Retail Prices API answers a question the two OpenRouter columns leave open about the cost of a reference image. For global consumption meters in East US 2, Azure lists text input, image input and image output at $5, $8 and $34 per million tokens for precision. Flash costs $1.75, $2.50 and $10 on the same three meters. The region’s separate data-zone meters have different prices. Azure Retail Prices API, East US 2 image meters

For an illustrative Azure edit metering 40 text-input tokens, 1,024 image-input tokens and 1,024 image-output tokens, the precision total would be $0.043208. Output alone accounts for $0.034816. On Flash the same three quantities total $0.01287. These are token-count assumptions. Use the metered input and output for your references and requested dimensions when putting a price on an actual brief.

The model card sets a ceiling of 4,194,304 output pixels, equivalent to 2,048 × 2,048. That square contains four times as many pixels as the leaderboard’s example. Azure’s public Microsoft model pricing page did not show a MAI-Image-2.7 price by resolution on November 19. A fourfold increase in pixels alone cannot supply the missing invoice price. For a campaign that needs the larger image, obtain the actual metered cost at the delivery dimensions. Azure’s public Microsoft model price table · How image API price units compare

03 — Worked budgetA catalogue of 100 products

Suppose a team needs four candidate compositions and one final image for each product. Use Microsoft’s rounded estimates for this illustration, with the same assumed square-image cost for every output. Before input charges, retries, storage and human review, 500 Flash images cost about $5. Five hundred precision images cost about $17.50. A mixed run of 400 Flash concepts and 100 precision finals costs about $7.50. These are estimates for generated outputs, with no assumed success rate or guarantee that an edit costs the same as a fresh generation.

Concepts
Four directions for each product
$4

400 Flash outputs at the rounded one-cent rate. The art director selects the composition to refine.

400 outputs
Finals
One precision image per product
$3.50

100 precision outputs at 3.5 cents each. Only the accepted composition reaches this pass.

100 outputs
Mixed total
A $2.50 premium over all Flash
$7.50

The premium buys a more precise attempt on the 100 important outputs. Reviewer time determines whether that trade paid off.

500 outputs

At an illustrative loaded review cost of $30 an hour, five minutes of a designer’s time costs $2.50. Saving five minutes across the entire set of 100 finals would cover the mixed route’s premium over all Flash. Conversely, if every final still needs the same manual packaging repair, the upgrade has bought little. Count variants requested, variants returned and images accepted. A report that only counts the 100 files delivered will conceal the 400 discarded concepts and any extra repair passes.

04 — Blind votesHow big is a four-point lead?

Microsoft cited the November 16 standings of AA-Image-Editing v2.1 in its launch announcement. The precision model recorded 1,215 Elo across 5,478 votes, four points ahead of GPT Image 2.6 Sunburst at max. The overall board still showed those scores on November 19. Artificial Analysis gives the same source image and edit instruction to competing models and asks evaluators to choose a result without knowing which model produced it. Elo expresses relative preference across those comparisons.

The published 95% interval for MAI-Image-2.7 spanned 1,201–1,229. Sunburst’s interval spanned 1,198–1,224 around its 1,211 estimate. Both intervals include both point estimates. Flash held seventh place with 1,179 Elo from 5,891 votes. That is evidence for putting the two Microsoft variants into different trials. The top row describes the precision model, and the board supplies no acceptance percentage for your brand or product. Microsoft’s separate top-two claim on Arena concerns the lab’s standing in generation and editing. Artificial Analysis image editing leaderboard and price definition

A blind reviewer can prefer a more attractive reflection without knowing that the product team rejected that reflection last quarter. Give your reviewer the approved reference, exact pack copy and intended placement. Record visual preference separately from a hard failure such as a missing accent in a product name. A model can win the first judgment and lose the second on the same image. GPT Image 2.6 and other editing models compared

05 — Usage rightsWhere can a team use client work?

MAI Labs and Canvas. The Labs Terms of Use limit access to personal use, and Microsoft’s model page explicitly excludes work and school activities, including use of the generated content. A small unpaid mock-up made for a client pitch is still work. Using a work or school account to sign in does not change the restriction. Keep that route for personal experiments with materials suitable for the preview. MAI Labs Terms of Use · Microsoft’s model and access notice

Microsoft Foundry. Microsoft Learn lists both models in public preview as Global Standard deployments. This is a route for an organization to evaluate the models under its Azure agreement. Microsoft’s supplemental terms say generative AI previews are intended for evaluation and caution against use in a live production environment. Before putting an automated image service on a customer’s site, check the specific deployment’s current status, the applicable Azure agreement and any model-specific terms. Azure supplemental preview terms · Microsoft Azure Product Terms

Partner APIs. Microsoft names fal, Genspark, OpenRouter and Vercel as launch partners. On fal, the precision and Flash editing API pages both carry a commercial-use label. The family FAQ expressly allows commercial projects using images generated through its API. Its API Services terms require customers to respect third-party rights and give no guarantee of originality. Those conditions matter when a brief includes licensed packaging artwork, a recognizable person or a competing brand. An API partner’s commercial label applies on its own service. fal’s precision editing endpoint · Flash editing endpoint · fal API Services terms

06 — Asset deliveryInspect the edges at size

Foundry documents separate image generation and image editing endpoints. An edit accepts an instruction and up to five JPEG or PNG references. The service returns PNG data that the application decodes. Assign each reference a job. For example, the first file defines the approved bottle shape, the second supplies a background, and the third shows the desired light direction. If a brief simply says to combine the pictures, a visually attractive blend may put the second image’s colors into the product itself. Microsoft Foundry generation and edit specifications

For a cutout, ask for a transparent background. Put the exported asset over both a dark and a light rectangle and inspect the edge at its intended display size. A faint white fringe can vanish against a white canvas and appear the moment the design switches to charcoal. Check translucent parts separately. A glass bottle needs its body to transmit some of the new background while the printed label stays opaque. File dimensions and a PNG extension alone cannot establish that either condition is true.

For a sequence of edits, compare the final candidate with the untouched source as well as the previous turn. The first turn can move the handle slightly, the next can shorten its shadow and the third can introduce an extra stitch. Each difference looks minor against its immediate predecessor. The original side by side with the final makes the accumulated drift visible. Keep the approved label or logo available as a separate layer if the final layout demands exact artwork. Seedream editing guide · MiniMax H3-Image reference and licence limits

07 — SelectionDecide with finished assets

Choose one recent brief with a subjective goal and one with an exact deliverable. A new set of seasonal backgrounds serves the first purpose. A photographed product with an approved bilingual label serves the second. Give both models the same inputs, dimensions and instructions. Have a reviewer mark the requested change, preserved features and time spent repairing the chosen candidate. If the work is destined for a specific partner’s API, run the trial there so its file limits, billing and permissions are represented.

Lots of exploratory choices
Start with Flash. Track how many distinct directions the designer would actually show and how long it takes to get the first usable set.
Flash
One detail must survive every edit
Try the precision model against the original. Include a packaging or text example that is easy to check and costly to repair.
MAI-Image-2.7
The asset is for a paid brief
Choose the service under which the work is permitted. Use a documented commercial endpoint for deliverables and evaluate Azure preview capabilities under the applicable terms.
Commercial route
Decision

Spend the extra cents where they save an edit

Flash and the precision model have different jobs when a designer is paid to deliver the final image. Buy inexpensive variations while the composition can change. Pay for fidelity once the product, label and placement are fixed. Price the route that can actually deliver that work and keep reviewer minutes on the same estimate as generated outputs. For help comparing creative workflows, see our AI transformation services.

Digital Applied

Price a creative brief from input to approval

Compare model charges and retouching time on the images your team ships.

Image workflowsCreative evaluationAPI costs
A representative brief

What to bring

  • →An approved source image
  • →The requested edits
  • →A final asset with known review time
Answers

Frequently asked questions

The November 19 Artificial Analysis editing board placed the precision model first at 1,215 Elo. Flash was seventh at 1,179. The board ranks preferences for instructed edits to existing images. A lab ranking on a separate leaderboard answers a different question.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Microsoft MAI Realtime 1: $0.01 Audio Minutes Plus Reasoning

Price Microsoft’s MAI Realtime 1 with its separate reasoning fees, compare voice API billing units and see what teams can test in the limited Foundry preview.

November 19, 2026 · 8 minRead
AI Development

ChatGPT Images 2.5: Flare, Sunburst and API Pricing

ChatGPT Images 2.5 adds new creative controls and two API models. Compare Flare and Sunburst, verified token prices, and a practical rollout plan.

September 8, 2026 · 8 minRead
AI Development

Gemini 4 Neon vs GPT-6.2 Sol vs Claude Sonnet 5.5 Pricing

Compare direct API prices, the 272K-token cutoff, cache costs and independent coding scores. See the bill for 100 jobs on Gemini 4 Neon, GPT-6.2 Sol and Sonnet.

November 19, 2026 · 9 minRead
AI Development

Microsoft MAI Code 2 and Flash: Price, Quality, Access

Compare MAI Code 2 and Flash in GitHub Copilot, including who gets each model, what Microsoft measured, and how token prices change on May 17, 2027.

November 19, 2026 · 8 minRead
AI Development

An Open Model That Makes Editable, Layered Designs

Ming-Image-0.1-Design makes UI, poster and infographic images with transparent backgrounds; its Layer variant splits a flat design into editable layers. MIT.

September 25, 2026 · 7 minRead
AI Development

Google Intelligent Eyewear: Gemini AI Glasses Fall 2026

Google announces Gemini-powered smart glasses with Samsung, Gentle Monster, and Warby Parker at I/O. Audio glasses ship fall 2026; display tier TBD.

May 20, 2026 · 18 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source