cd /news/artificial-intelligence/microsoft-s-mai-image-2-7-tops-editi… · home › topics › artificial-intelligence › article
[ARTICLE · art-171354] src=alphasignal.ai ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Microsoft's MAI-Image-2.7 Tops Editing Charts at One-Sixth the Cost

Microsoft released MAI-Image-2.7 and MAI-Image-2.7-Flash in public preview on Microsoft Foundry, with the full model ranking first on Artificial Analysis's AA-Image-Editing v2.1 leaderboard at a normalized $34.82 per 1,000 images — roughly one-sixth the price of the four models ranked immediately below it. The Flash tier costs an estimated $10.24 per 1,000 images, down 47% from MAI-Image-2.6-Flash at $19.46, and both models are available through OpenRouter, Microsoft Foundry, MAI Labs, fal, Genspark and Vercel. MAI-Image-2.7 leads four of the benchmark's seven editing actions — Enhancement and Restoration, Reasoning-Based Edits, Scene and Style, and Text or Symbol Editing — but shows its largest deficit in Object-Level Edit and ranks fourth on text-to-image.

read5 min views1 publishedNov 17, 2026
Microsoft's MAI-Image-2.7 Tops Editing Charts at One-Sixth the Cost
Image: Alphasignal (auto-discovered)

Microsoft's new MAI-Image-2.7 and 2.7-Flash take #1 on image editing at roughly one-sixth the cost of rival frontier models.

  • Microsoft AI launched MAI-Image-2.7 and MAI-Image-2.7-Flash in public preview on Foundry.
  • MAI-Image-2.7 takes #1 on AA-Image-Editing v2.1, beating GPT Image 2.6 Sunburst (max).
  • Priced at $34.82 per 1k images, roughly one-sixth the cost of rival frontier editors.
  • Flash tier drops to $10.24 per 1k images, down from $19.46 for MAI-Image-2.6-Flash.
  • Available via OpenRouter , Microsoft Foundry, MAI Labs, fal, Genspark and Vercel.
  • Wins on text edits, scene and style, restoration; weaker on object-level edits and lighting.

Microsoft has released MAI-Image-2.7 and its faster Flash variant in public preview, moving its in-house image stack from a playground demo to callable services. On Artificial Analysis’s editing leaderboard, the full model ranks first on AA-Image-Editing v2.1. Its normalized price is roughly one-sixth of the prices shown for the four models immediately below it.

Microsoft’s prior October preview offered no API or pricing, which prevented developers from integrating it into products. The public preview adds metered access through Microsoft Foundry and several third-party platforms.

The cost curve drops #

Artificial Analysis reports full-model rates of $5 per million text input tokens, $8 per million image input tokens, and $34 per million image output tokens. Its normalization converts those rates to an estimated $34.82 per 1,000 images. MAI-Image-2.7-Flash costs an estimated $10.24 per 1,000 images, down 47% from MAI-Image-2.6-Flash at $19.46.

Artificial Analysis leaderboard snapshot
Model Editing rank Text-to-image rank Normalized cost per 1,000 images
--- --- --- ---
MAI-Image-2.7 1 4 $34.82
MAI-Image-2.7-Flash 7 11 $10.24

At 100,000 images, those normalized rates translate to $1,024 for Flash, $3,482 for the full model, and $1,946 for the previous Flash release. Actual invoices will vary with input-image tokens, output dimensions, image count, and provider fees.

Scene-wide edits lead #

Across the benchmark’s seven editing actions, MAI-Image-2.7 ranks first in four: Enhancement and Restoration, Reasoning-Based Edits, Scene and Style, and Text or Symbol Editing. It also leads the Architecture and Real Estate, Consumer, Frontier, and Marketing and Advertising use-case categories.

  • Scene and Style: restyling or relighting a scene while preserving the subject’s identity and composition.
  • Text or Symbol Editing: replacing lettering while matching the original material, perspective, and lighting.
  • Enhancement and Restoration: repairing degraded assets, increasing clarity, and recovering visual detail.
  • Reasoning-Based Edits: changing objects or environments while maintaining plausible geometry, shadows, and relationships.

Flash ranks seventh overall for editing, with third-place results in Scene and Style, Frontier, and Animation and Gaming. It ranks fourth in Enhancement and Restoration. That profile makes it a candidate for high-volume drafts, batch variations, and first-pass edits, with the full model available for requests that miss quality thresholds.

Object edits expose the gap #

The category breakdown shows MAI-Image-2.7’s largest editing deficit in Object-Level Edit. Lighting is its weakest text-to-image category, and its overall text-to-image rank is fourth. The benchmark’s higher-ranked GPT Image and Seedream entries remain stronger candidates for workloads centered on generating new images from prompts. Flash ranks eleventh for text-to-image, reinforcing its focus on cost-efficient editing.

Leaderboard positions summarize aggregate benchmark performance and can shift as models or evaluations change. The cited data covers quality rankings and normalized prices; production latency, rate limits, failure rates, safety behavior, and provider-specific image processing require separate measurement.

The demo becomes an API #

First-party access is available through Microsoft Foundry and MAI Labs, whose Canvas experience runs on 2.7-Flash. OpenRouter, fal, Genspark, and Vercel also offer listings or integrations. The OpenRouter listing provides one route for testing the model through an OpenAI-compatible endpoint.

Each editing request can include text and as many as five reference images. Multiple references allow a single request to carry a target asset alongside style, lighting, composition, or brand references. Prompts should identify each image’s role explicitly because ordering alone may leave the intended relationship ambiguous.

OpenRouter’s chat endpoint can send the target and style reference as data URLs and request an image response:

import base64
import mimetypes
import os
from pathlib import Path

import requests

def data_url(path):
    mime = mimetypes.guess_type(path)[0] or "application/octet-stream"
    encoded = base64.b64encode(Path(path).read_bytes()).decode()
    return f"data:{mime};base64,{encoded}"

payload = {
    "model": "microsoft/mai-image-2.7",
    "modalities": ["image", "text"],
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": (
                        "Use the first image as the target. Use the second "
                        "only as a lighting and style reference. Place the "
                        "product in a sunlit Scandinavian kitchen."
                    ),
                },
                {
                    "type": "image_url",
                    "image_url": {"url": data_url("product.png")},
                },
                {
                    "type": "image_url",
                    "image_url": {"url": data_url("style_ref.png")},
                },
            ],
        }
    ],
}

response = requests.post(
    "https://openrouter.ai/api/v1/chat/completions",
    headers={
        "Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}",
        "Content-Type": "application/json",
    },
    json=payload,
    timeout=120,
)
response.raise_for_status()

message = response.json()["choices"][0]["message"]
images = message.get("images", [])

if not images:
    raise RuntimeError("The response contained no image.")

image_url = images[0]["image_url"]["url"]

if image_url.startswith("data:"):
    encoded = image_url.split(",", 1)[1]
    Path("edited.png").write_bytes(base64.b64decode(encoded))
else:
    print(image_url)

Preview status makes endpoint behavior, response formats, quotas, regional availability, model revisions, service-level commitments, and data-handling terms provider-specific. Applications should follow the selected provider’s current documentation and avoid depending on unversioned response fields without validation.

Build the production eval #

A production evaluation should route the same representative assets through Flash, full 2.7, and an existing baseline while recording both quality and operating data:

  • Instruction fidelity and preservation of untouched regions
  • Identity, product geometry, logos, and embedded text
  • Consistency across repeated requests and reference-image combinations
  • Latency, timeout rate, retry rate, and provider limits
  • Actual billed cost at required resolutions and output counts
  • Moderation behavior, retention terms, licensing, and regional processing

A two-stage routing pattern can send bulk variants to Flash and retry failed acceptance checks on the full model. At the normalized rate of about one cent per Flash image, that design supports workloads such as personalized advertising, catalog variations, architectural staging, and large-scale asset restoration while retaining a higher-quality path for difficult edits.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @microsoft 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/microsoft-s-mai-imag…] indexed:0 read:5min 2026-11-17 · —