TL;DR
MAI-Image-2.7-Preview is Microsoft's AI image-editing model, introduced on October 27, 2026. It focuses on controlled changes to existing images, including lighting, visual style, objects, and typography.
Microsoft announced MAI Image 2.7 at No. 3 on Arena Edit and No. 2 on Artificial Analysis. In the October 28, 2026 snapshot reviewed for this article, it scored 1,193 Elo on Artificial Analysis and ranked second. For designers, its value lies in making substantial edits while preserving elements that already work.
Developers can evaluate the model through public testing platforms, but Microsoft has not announced a public 2.7 API or pricing as of October 28, 2026.
Key Takeaways
- Released October 27, 2026: Microsoft introduced MAI-Image-2.7-Preview as an image-editing-focused successor to MAI-Image-2.6.
- 1,193 Elo on Artificial Analysis: The model ranks second on AA-Image-Editing v2.1, six points behind GPT Image 2.5 Sunburst.
- Broader editing gains: Microsoft reports improvements across all 10 evaluated use cases and seven editing actions, rather than a gain in only one type of edit.
- Advanced editing capabilities: The model targets scene restyling, relighting, text and symbol editing, reasoning-based modifications, and image restoration.
- Evaluate before integrating: Test MAI-Image-2.7-Preview against your own preservation requirements; use an available API when your application needs deployment today.
What Is MAI-Image-2.7-Preview?
MAI-Image-2.7-Preview is a proprietary image-editing model developed by Microsoft AI. It belongs to Microsoft's MAI Image family, which includes earlier models designed for image generation, image editing, and production-oriented creative workflows.
Instead of replacing the entire image, the model is designed to make targeted edits while preserving the original scene’s visual coherence.
For example, a designer might place a product in a new environment while preserving its packaging, logo, and proportions. A photographer might want to alter lighting without distorting the subject. A marketing team might need to replace text on an existing advertisement while preserving its layout.
MAI-Image-2.7-Preview specifications
Microsoft has not disclosed the preview's architecture, parameter count, maximum resolution, or reference-image limit in its launch announcement. Earlier MAI specifications should not be treated as confirmed 2.7 limits.
What Can MAI-Image-2.7-Preview Do?
Microsoft says the model reaches the category frontier in four areas: enhancement and restoration, reasoning-based edits, scene and style editing, and text or symbol editing.
Scene and style editing
MAI-Image-2.7-Preview can modify the overall appearance of an existing image, including its environment, artistic style, lighting, and mood.
Consider a product photograph taken against a plain studio background. A designer may want to transform the surrounding environment into a premium outdoor campaign scene while preserving the product itself.
A successful result requires more than changing the background. The lighting direction, reflections, shadows, perspective, and colors must remain consistent.
Microsoft reports a 24-point Elo lead in scene and style editing over the next-best model in that category. That result makes background changes and relighting useful starting points for your own evaluation.
Text and symbol editing
AI image models often struggle with modifying text that is already embedded in an image.
Replacing a label can accidentally change its typography, positioning, perspective, or surrounding design. Even when the replacement words are correct, the edited region may look artificially pasted onto the original.
MAI-Image-2.7-Preview focuses on changing text and symbols while keeping them visually integrated.
Likely applications include product packaging, advertisements, promotional posters, presentation graphics, and UI mockups.
For businesses creating multilingual marketing assets, this could be particularly valuable. However, the model's reliability across languages and complex typography still requires task-specific testing.
Reasoning-based image editing
Some image edits require an understanding of how objects interact with their surroundings.
For example, moving a reflective object into a different environment can require coordinated changes to lighting, shadows, and nearby surfaces.
An editing model may technically follow the instruction while producing an inconsistent scene.
MAI-Image-2.7-Preview is designed to handle edits involving relationships between objects, lighting, and environments.
Image enhancement and restoration
Typical workflows include restoring old photographs, enhancing low-quality product images, refining creative assets, and correcting visual imperfections.
AI restoration can invent plausible details. For archival or evidence-based work, keep the original and clearly identify generated changes.
These are capability categories rather than guarantees of perfect results. Actual editing accuracy depends on the source image, prompt complexity, and review criteria.
How Does MAI-Image-2.7-Preview Perform on Benchmarks?
MAI-Image-2.7-Preview has appeared on two major image-editing evaluation platforms: Artificial Analysis and Arena.
Both use preference-based evaluation approaches, but they have different datasets, scoring scales, and voting methodologies.
Their Elo scores should be compared within each leaderboard, not directly across platforms.
Artificial Analysis: AA-Image-Editing v2.1
Artificial Analysis evaluates how well image models modify an input image in response to text instructions.
AA-Image-Editing v2.1 ranks models using pairwise human-preference votes on image-editing tasks. These are external platform results, rather than an in-house Microsoft benchmark.
The October 28, 2026 snapshot shows:
The table preserves the published model settings, including the max setting for GPT Image 2.5. The ± values are the displayed 95% uncertainty intervals. The leaderboard does not provide a shared hardware or latency test environment, so these scores describe editing preference rather than processing speed.
MAI-Image-2.7-Preview scores 42 Elo above MAI-Image-2.6, 11 above Flare, and six below Sunburst on this leaderboard.
However, the confidence intervals overlap among the leading models. A six-point difference does not establish that Sunburst will consistently outperform MAI-Image-2.7-Preview on every editing task.
Arena Edit: Image editing
Arena also measures user preferences through blind comparisons.
The overall single-image editing scores visible on October 28, 2026 are shown below. The table retains Arena's preliminary labels where displayed.
The displayed scores place 2.7 Preview 44 Elo above 2.6 and 11 Elo above GPT-Image-2 (Medium).
For additional API baselines, developers can access GPT Image 2 and Grok Imagine Image 2.0 API. Medium in Arena's GPT Image 2 entry identifies an evaluation setting, not a separate API model name.
Arena results are live and may change over time; the figures below reflect the October 28 snapshot reviewed for this article.
Commercial design performance
Arena's Product, Branding & Commercial Design category is especially relevant when the output must work as an advertisement or product asset.
In Arena’s Product, Branding & Commercial Design category on October 28, 2026, MAI-Image-2.7-Preview ranked third at 1,499 Elo—55 points ahead of 2.6, which ranked eighth. With only 718 votes and a wider uncertainty interval, the result is an early signal to test against your own design requirements.
Both leaderboards show stronger editing performance than MAI-Image-2.6 and competitive results against leading models. Production speed, reliability, and cost still require separate API testing.
MAI-Image-2.7-Preview vs MAI-Image-2.6
MAI-Image-2.6 was previously made available to developers through Microsoft Foundry, with a Flash variant offered separately. That API milestone is distinct from the August release month recorded by Artificial Analysis.
MAI-Image-2.6 and its Flash variant introduced developer-facing capabilities such as multi-image reference editing, web grounding, dynamic aspect ratios, and output resolutions up to 1.5K.
The 2.7 preview shifts the immediate release emphasis toward editing performance.
| Dimension | MAI-Image-2.6 | MAI-Image-2.7-Preview |
|---|
Microsoft reports gains across all 10 evaluated use cases and seven editing actions. The practical question is whether those gains improve the edits your team performs most often.
The previous model remains relevant for developers who need a documented image-generation API and established production integrations.
How to Access and Try MAI Image 2.7
Microsoft identifies MAI Playground, Arena, and Artificial Analysis as routes for exploring the preview. Start with the Playground for your own image, and use the evaluation platforms to compare preferences and scores.
The MAI Playground is Microsoft's direct experimentation environment for MAI models.
When the preview is available to your account and region, the intended workflow is to select the editing model, provide an existing image, and describe the desired changes.
A useful first experiment is an edit with one major change and several explicit preservation requirements.
Example prompt:
Replace the background of this product photograph with a modern luxury retail interior. Preserve the product's exact proportions, logo, packaging text, and camera angle. Adjust the lighting and shadows to match the new environment. Do not introduce additional text.
A repeatable testing method
To evaluate the model meaningfully, use the same source images and instructions across competing editors.
For an initial screening, assemble 20 to 50 representative tasks covering product changes, typography, relighting, restoration, and multi-object compositions. This is a practical starting set, not a statistically validated sample-size requirement.
Review outputs against four dimensions:
- Instruction accuracy: Did the model make the requested changes?
- Preservation: Did untouched elements remain consistent?
- Visual realism: Do the lighting, perspective, and composition look natural?
- Output acceptance: Can you use the result without further correction?
For a production-oriented assessment, latency and cost should also be recorded once suitable API access becomes available.
Is MAI-Image-2.7-Preview API Available, and What Does It Cost?
As of October 28, 2026, Microsoft has not announced a public API endpoint, pricing, or general-availability date for MAI Image 2.7. Artificial Analysis labels the preview No API available.
For production planning, the unresolved details include the deployment model ID, image limits, rate limits, billing, and preview-specific usage terms.
Microsoft says text-to-image generation is coming soon. If your application needs both generation and editing now, choose a documented API and treat 2.7 as a separate evaluation candidate.
Image Editing with CometAPI
For developers who need image editing in production today, CometAPI offers API access to several established alternatives.
The GPT Image 2.5 API in CometAPI supports image generation and editing through OpenAI-compatible endpoints.
This allows you to validate the editing workflow now and evaluate MAI Image 2.7 separately when Microsoft provides a supported developer interface.
The example below calls GPT Image 2.5 Sunburst. It gives you a working integration pattern for an available alternative; it does not call MAI-Image-2.7-Preview.
Example: Editing an existing image with CometAPI
Run this request in Bash with a valid CometAPI key and product.png in the current directory. It saves the JSON response to edit-response.json.
export COMETAPI_KEY="YOUR_COMETAPI_KEY"
curl --fail-with-body \
"https://api.cometapi.com/v1/images/edits" \
-H "Authorization: Bearer $COMETAPI_KEY" \
-F "model=gpt-image-2.5-sunburst" \
-F "image[]=@product.png" \
-F "prompt=Preserve the product, logo, and camera angle. Replace only the background with a premium studio environment. Match the original lighting naturally." \
-F "quality=high" \
-F "output_format=png" \
--output edit-response.json
According to the CometAPI response format, the edited image is returned as a Base64 value in data[0].b64_json. Decode that Base64 value into a PNG file; the JSON response itself is not an image. If cURL reports an HTTP error, inspect the response body before retrying.
Production evaluation metrics
A useful measure of production efficiency is cost per accepted image, not merely cost per request.
For example, suppose an editing service produces 100 images at an average total cost of $0.10 each.
If only 70 pass quality review, then:
Cost per accepted image = $10 / 70 = approximately $0.143.
If improved editing quality raises the acceptance rate to 90% at the same per-request cost, the effective cost falls to approximately $0.111 per accepted image.
This hypothetical example isolates the effect of acceptance rate; the $0.10 input assumption is not a quote for any model.
The example shows why editing accuracy can matter as much as the headline API price.
FAQ
How should I handle a distorted logo in a MAI-Image-2.7-Preview edit?
Return to the original image, narrow the requested change, and state explicitly that the logo and packaging text must remain unchanged. Compare the relevant crop against the original at full resolution. If the model still changes brand-critical details, preserve those elements through a separate compositing step or use a conventional editor for that region. A visually attractive result is not sufficient when the logo must be exact.
Can repeated MAI-Image-2.7-Preview edits accumulate errors?
Repeated generative edits can introduce drift, even when each individual result looks plausible. Keep the untouched source and save versions after each accepted edit. For a new variation, compare starting from the original with continuing from the last output; choose the path that preserves the details your project requires. Do not assume that an earlier successful edit guarantees consistency through a longer sequence.
How can I compare MAI-Image-2.7-Preview results fairly?
Hide the model names during review and randomize the output order. Agree on rejection criteria before reviewers see the images, and record failures as well as preferred outputs. If you allow retries, give every model the same retry budget and include all attempts in the cost calculation. This reduces the chance of favoring a model because one unusually strong result was selected.
Does MAI-Image-2.7-Preview restoration preserve historical truth?
A restored image may contain plausible details that were not visible in the source. Keep the original alongside the edited version and distinguish cosmetic improvement from factual reconstruction. For an archive or documentary project, record the requested changes and identify generated details rather than treating the restored image as new evidence about the original scene.
Conclusion
Evaluate MAI-Image-2.7-Preview if your workflow depends on making substantial edits while preserving key visual details. Use representative source images to measure preservation quality and output acceptance. If you need an API integration today, choose an available model and revisit 2.7 when Microsoft publishes its developer access and pricing.