What Is Microsoft-Decision-1.1? Multimodal Routing, Calibration, and Pricing
SUMMARY
Microsoft-Decision-1.1 is Microsoft's specialized decision model, generally available since November 17, 2026, built to select constrained software actions rather than generating open prose. Version 1.1 adds multimodal support for photos and screenshots alongside calibrated confidence scores so applications can act, split confidence, or abstain when inputs are unclear. Developers access it through Microsoft Foundry using input-token billing, as well as third-party gateways, to power high-volume routing and triage. Because a decision model only classifies inputs and picks software branches, it does not produce finished promotional visuals. An independent business still supplies product photos to a creative service like Stillmore, whose team handles production and pre-delivery checks before delivering ready-to-post images, videos, and captions for the merchant to publish.

Contents
What Is Microsoft-Decision-1.1?
Microsoft-Decision-1.1 is a task-specific model engineered to pick from a developer's permitted software actions and report how certain it is about each choice. Unlike conversational chatbots rewarded for writing fluent paragraphs, a decision model operates inside strict application boundaries where structured routing, tagging, or verification matters more than prose style.
According to Microsoft's GA announcement, the model pairs a custom inference engine with Gemma 4 26B A4B and SFT plus reinforcement learning for abstention and calibrated scoring; it reached 68 on the AA Decision Index v2.1, achieved the highest average accuracy on Microsoft's separate 102 blind benchmarks, served over 4,000 preview customers, and lists Foundry input pricing at $0.042 Global, $0.046 US, and $0.05 EU/APAC per million tokens with free output. Separate Ignite 2026 coverage lists Microsoft-Decision-1.1 as 50x faster and 79x cheaper than GPT-Sol 6.1.
Consider an illustrative customer support queue receiving an order-issue ticket with an attached phone photo of a damaged parcel. Instead of drafting a polite email reply, the model evaluates the developer's permitted label set, such as approve_replacement, request_clearer_photo, route_carrier_claim, or abstain, and returns a structured choice with a confidence score that backend software can route immediately.
| Task Category | Typical Input Supplied | Role of a Decision Model | Remaining Work Outside the Model |
|---|---|---|---|
| Support Ticket Triage | Order text paired with a damaged-item snapshot | Selects queue tag or requests clearer image | Resolving refunds, carrier claims, or customer replies |
| Catalog Quality Gate | Merchant listing title and storefront photo | Flags blurry uploads or mismatched categories | Re-shooting items, editing listings, or approving inventory |
| Incident Knowledge Lookup | System alerts, logs, and dashboard screenshots | Matches incoming signals to relevant runbook routes | Investigating root cause and applying production fixes |
| Retail Promo Intake | Everyday product photos and promotion notes | Could pre-sort angles or flag missing info if integrated | Creative direction, compositing, copywriting, and delivery checks |
Why Microsoft-Decision-1.1 Matters for Software and Media Workflows
Microsoft-Decision-1.1 targets three recurring engineering bottlenecks, high inference cost, slow sequential token decoding, and overconfident model guessing, while adding visual intake for teams handling photos and screenshots. Embedding model calls deep inside high-volume software stalls when every simple branching check costs as much and takes as long as writing a multi-paragraph answer.
Standard generative models are trained to complete tasks and sound persuasive even when evidence is weak, which makes automated software execution risky when an image is cropped or a message lacks context. Adding direct photo and screenshot evaluation inside a low-latency decision call lets engineering teams filter incomplete visual submissions before downstream workflows act on bad assumptions.
Suppose an operations team handles thousands of daily merchant submissions pairing a short product description with a counter snapshot. Asking a large conversational model to write free-form explanations for every item wastes compute and risks silent misclassification when glare hides a label. A dedicated visual router either sorts clear submissions into the right catalog queue or pauses ambiguous frames for human inspection.
How Microsoft-Decision-1.1 Works: Calibration, Abstention, and Inference
Microsoft-Decision-1.1 combines a non-token-by-token inference architecture with calibration training so applications can measure uncertainty before executing code. Because decision calls often accept long context or visual payloads and return compact choices, optimizing the runtime around branching evaluation rather than sequential prose generation reduces runtime delay and hardware overhead.
Training a decision model to know what it does not know changes how software engineers write fallback logic. Instead of forcing a single confident guess on borderline inputs, Microsoft's calibration training lets the model report strong certainty on clear cases, split confidence when multiple options are valid, or abstain altogether when required visual or text evidence is missing.
In a hypothetical deployment scenario, an engineering team replaces a brittle binary filter with four threshold rules tailored to operational risk. Rather than letting software execute every top label blindly, the backend checks both the winning action and score separation before deciding whether automation proceeds or pauses for an operator.
| Confidence State | Underlying Application Signal | Recommended Routing Rule | Human Review Trigger |
|---|---|---|---|
| Single High Confidence | One permitted action clears the strict cutoff | Execute downstream routing or tag automatically | Periodic spot audit only |
| Split Confidence | Two or more valid options share mid-range scores | Attach candidate tags and pause auto-execution | Route to queue for rapid human tie-break |
| Incomplete or Ambiguous Input | Model abstains due to missing text or obscured photo | Reject silent automation and request clearer material | Notify sender or specialist with specific gap |
| High-Risk Action Boundary | High score on billing, compliance, or irreversible state | Enforce second-stage rule check before execution | Require explicit operator sign-off |
Practical Applications: From Support Triage to Photo Intake Routing
In production environments, Microsoft-Decision-1.1 handles high-frequency routing, effort scaling, and evaluation judging inside software, while human creative specialists remain essential whenever raw visual material must become finished promotional work. Sorting an asset into a folder is a classification step; designing, compositing, and verifying brand media is a production step.
Microsoft's GA announcement reports using the model internally for request routing across Copilot experiences, dynamic reasoning control in Cowork, evaluation judging, and on-call incident lookup. By contrast, Inside Stillmore: Our Creative Production Workflow details how finished marketing content relies on photography and live-action direction, post-production and VFX work in Photoshop and After Effects, and an AI production team supported by software agents to inspect product shape, color, text, motion, and composition before delivery.
To see where automated routing ends and creative production begins, imagine an independent neighborhood café uploading 15 phone snapshots of a seasonal layered drink for next month's social calendar. A classifier could help tag close-up cup shots versus wide counter angles, and Stillmore's team may evaluate and use models such as Microsoft-Decision-1.1 when they suit a brief, combining them with editing and creative review to pursue stronger finished work. Turning those snapshots into customer-ready posts still belongs to Stillmore: its specialists pick usable angles, preserve the drink's layers and cup label across motion treatments, verify offer wording provided by the shop, and deliver finished images, videos, and captions for the owner to publish.
Microsoft-Decision-1.1 vs. General-Purpose LLMs and Creative Production Workflows
Microsoft-Decision-1.1 is not a direct substitute for either a general-purpose LLM or a creative production agency because each operates on a different type of deliverable. Comparing them side by side clarifies what software teams gain from calibrated routing and what work remains untouched after a model returns a label.
A decision model selects constrained software actions rather than writing articles, synthesizing multi-paragraph arguments, or creating visual assets. General-purpose LLMs remain appropriate when an application needs open prose, code synthesis, or flexible reasoning, even though they carry higher per-call overhead and require tighter prompting to limit overconfident answers. Neither model class replaces a creative studio's pre-delivery inspection of physical product geometry, accurate label text, and brand-appropriate pacing, as outlined in Stillmore's workflow guide.
For an illustrative shop scenario, sending a sneaker photo into a decision router can tell an internal workflow whether the shoe is centered, whether the outsole is visible, or whether a blurry frame should be flagged. That structured output will never place the sneaker onto a styled seasonal surface, preserve exact stitching under camera motion, or deliver a publish-ready short video with matched social copy.
| Dimension | Microsoft-Decision-1.1 | General-Purpose LLM | Creative Production Service (e.g., Stillmore) |
|---|---|---|---|
| Primary Output | Permitted action label and calibrated confidence score | Open-ended text, code, or conversational responses | Checked promotional images, short videos, and captions |
| Role of Image Inputs | Classifies photos and screenshots to guide branching logic | Describes, discusses, or analyzes visual context | Starting reference material refined into finished brand assets |
| Handling Uncertainty | Trained to split confidence or abstain on gaps | Tends to complete tasks smoothly unless tightly constrained | Team requests missing angles or separate offer wording |
| Remaining Work for User | Build API integration, thresholds, and downstream actions | Verify factual accuracy, format output, handle edge cases | Choose when and where to publish finished assets |
When Microsoft-Decision-1.1 Fits, and When It Does Not
Microsoft-Decision-1.1 fits engineering teams running high-volume triage, routing, evaluation, or visual intake filters, and it does not fit open-ended creative drafting or turnkey business marketing. Knowing which side of that line your problem sits on prevents paying for engineering complexity, or overpaying for frontier chat models, where neither is needed.
Because Microsoft's published Foundry billing charges only on input context rather than returned labels, teams feeding large batches of logs, support threads, or screenshots get predictable unit economics. Exact per-token pricing on third-party gateways is omitted here because those partner rates were not published in the November 17 announcement.
In an illustrative budget scenario, a helpdesk platform classifying millions of mixed text-and-screenshot tickets each month benefits directly from input-based routing and calibrated abstention. Conversely, a local bakery wanting polished promotional videos and captions every month gains nothing from wiring up model endpoints, maintaining schema constraints, or hiring editors separately to act on routing labels.
How to Get Started with Microsoft-Decision-1.1
Developers can access Microsoft-Decision-1.1 across official cloud and gateway channels once they define a closed list of permitted actions and calibrated fallback thresholds. Setup success depends less on conversational prompt styling and more on crisp boundaries between automated execution and human escalation.
I reviewed the official deployment listings on Microsoft Foundry, OpenRouter, and Vercel AI Gateway. Because this overview is based on official documentation review rather than hands-on API benchmarking, treat the four-step setup sequence below as an editorial implementation recommendation rather than an observed production test:
- Map every permitted outcome your workflow supports, ensuring an explicit
abstainorescalatebranch exists for incomplete data. - Select your deployment platform, choosing Microsoft Foundry for regional data zones and input-based billing, or OpenRouter and Vercel AI Gateway if your stack already standardizes there.
- Pair text criteria with explicit photo or screenshot checks so the model evaluates specific visual conditions before routing.
- Calibrate tiered confidence cutoffs in staging using ambiguous real-world edge cases before allowing automated execution in production.
In a hypothetical onboarding scenario, a SaaS billing team designs a receipt-verification route with four permitted choices: approve_receipt, request_uncropped_photo, flag_mismatch, and abstain. Suppose during hypothetical staging checks on 500 historical screenshots, cropped mobile receipts tend to score around 0.55 to 0.72 confidence; the engineers could set an auto-approval cutoff at 0.90, route scores between 0.55 and 0.89 to a human queue, and trigger a re-upload message whenever the model selects abstain.
Conclusion
Microsoft-Decision-1.1 establishes a clear divide between automated software routing and finished creative production. For engineers building responsive applications, developers pair model confidence outputs with measured staging evaluations to set explicit thresholds for when software executes automatically versus when it escalates, recognizing that probabilistic scores are never guaranteed correct on every edge case. For independent businesses focused on everyday marketing rather than software pipelines, the practical divide stays clear: a decision model sorts signals inside apps, while creative production services turn existing phone photos into checked, publish-ready images, videos, and captions.
Related reading
- Inside Stillmore: Our Creative Production Workflow
- What Is Stillmore? Content Creation for Small Businesses
Resources
| Source | Publisher / Platform | Direct Link | Scope Covered |
|---|---|---|---|
| Microsoft-Decision-1.1 GA Announcement | Microsoft Command Line | Microsoft-Decision-1.1 GA post | Official GA announcement and specifications |
| Microsoft Ignite 2026 Model Lineup | Microsoft AI | Ignite 2026 announcement | Ignite 2026 model lineup comparison |
| Microsoft Foundry Deployment | Microsoft | Microsoft Foundry listing | Microsoft Foundry deployment page |
| OpenRouter Model Listing | OpenRouter | OpenRouter model page | OpenRouter model availability page |
| Vercel AI Gateway Changelog | Vercel | Vercel AI Gateway changelog | Vercel AI Gateway release note |
| Stillmore Production Workflow | Stillmore | Inside Stillmore: Our Creative Production Workflow | Stillmore production teams and quality checks |
Frequently Asked Questions
Does Stillmore already use Microsoft-Decision-1.1 to sort client product photos?
No, Stillmore has not confirmed using Microsoft-Decision-1.1 in production, and our creative team currently picks client photos directly from each business's Library. When new models release, our specialists may evaluate and use options such as Microsoft-Decision-1.1 when they suit a brief, pairing them with editing and human review to pursue stronger finished work. Merchants never need to track internal model evaluations because Stillmore's team handles photo selection, creative direction, and pre-delivery quality checks before anything reaches your account, keeping your existing snapshots reusable across seasonal promotions and product reminders.
Can Microsoft-Decision-1.1 generate finished promotional images or videos like Stillmore?
No, Microsoft-Decision-1.1 outputs structured software routing rather than edited promotional images, short videos, or social captions. A decision classifier only triggers application branches behind the scenes, whereas Stillmore turns your everyday business photos into complete marketing assets ready for your social channels. Our photography, post-production, VFX, and AI production teams handle compositing, retouching, motion work, and copywriting, checking every visual detail against your reference photos so you receive finished posts ready to share instead of raw classification labels.
Who checks whether a phone photo shows enough product detail before production starts?
Both the merchant and Stillmore's team verify product clarity before production starts, beginning with a quick zoom check on your phone. Before uploading, enlarge the item to confirm its label, outline, and distinguishing features stay legible without heavy glare or cropped edges. During production, Stillmore checks those reference photos alongside the exact product name and offer wording you provide separately. Because editing cannot reliably establish hidden packaging or obscured sides, our team may request another angle whenever important details remain unclear.
How does Stillmore combine software workflows with human creative review?
Stillmore uses software agents to support workflow coordination across production stages while human specialists keep responsibility for creative judgment, editing, and inspection. Our AI production team selects professional generation tools suited to the concept alongside agentic workflow support. Meanwhile, our photography and post-production teams shape lighting, framing, Adobe Photoshop compositing, and Adobe After Effects motion graphics. Before delivery, reviewers check product shape, color, text accuracy, motion timing, and small-screen composition so the finished work matches the item you actually sell.
What happens when an uploaded business photo is ambiguous or missing offer details?
Stillmore does not guess promotion dates, product names, or hidden packaging details when an uploaded photo is ambiguous or incomplete. A snapshot shows how an item looks on your counter, yet it cannot tell our team when a weekend discount ends or which exact wording belongs on screen. Merchants supply accurate product names and promotion copy separately during submission, and upload replacement angles whenever labels are obscured or packaging changes so finished promotional assets reflect truthful business information.
Do small businesses need to configure Foundry or API thresholds to work with Stillmore?
No, small businesses working with Stillmore never configure cloud deployments, manage API keys, or tune model thresholds. Software engineers use decision endpoints to automate branching inside custom applications, whereas Stillmore operates as an advertising creative service rather than a developer toolkit. You simply upload existing product, storefront, or team photos, choose an Inspiration style or enable ongoing creation, and share offer details. Our team handles internal tool choices, editing software, and pre-delivery quality checks behind the scenes without adding technical work.
How do merchants choose between picking an Inspiration style and automatic creation?
Merchants switch between manual Inspiration selection and automatic creation inside My Creations, and both modes share the same Plan Credits. When launching a new item or seasonal offer, you can browse Inspiration styles, pick a preferred look, and submit your chosen photos yourself. During busy weeks, enabling "Stillmore creates for me" lets our team pick suitable Library photos and schedule ongoing creation and delivery. For example, requesting two manual videos on Growth leaves remaining monthly Credits for four team-planned videos.
How many photos are required in Stillmore Library before ongoing creation starts?
Stillmore requires at least 3 photos in your Library for Try it or 15 photos for Growth (monthly or annual) before automatic creation starts. Meeting that threshold gives our team enough varied angles to reuse your saved product, storefront, or team photos across ongoing content without monthly re-uploads. If your Library holds fewer photos than the automatic minimum, you can still browse Inspiration and submit manual requests with their required materials, uploading fresh snapshots whenever new items or angles appear.
Who approves and publishes finished videos, images, and captions?
Stillmore inspects and delivers finished videos, images, and captions, while each merchant keeps full control over approval and publishing. On Growth automatic mode, once at least 15 Library photos, payment, Credits, suitable materials, and automatic mode are ready, delivery normally runs one video every five days; manual requests also normally arrive within five days once required materials and payment are complete. Late-month uploads may not finish by month-end, and you decide when and where to post finished assets on your social channels.
How do Stillmore Credits compare to token-based billing on Microsoft Foundry?
Token-based cloud billing charges software teams by raw text or data volume, whereas Stillmore prices finished promotional deliverables using Plan Credits. On Stillmore, one finished video uses 3,000 Credits and one finished image uses 500 Credits, covering creative planning, editing, captions, and pre-delivery checks rather than raw model calls. Unused Plan Credits stay available and can accumulate up to two months of your plan's allocation, equal to 6,000 Credits on Try it or 36,000 on Growth, while purchased Add-on Credits never expire.
When should a business build with Microsoft-Decision-1.1 versus choosing Stillmore?
Choose Microsoft-Decision-1.1 when an engineering team needs automated software branching inside a custom application, and choose Stillmore when an independent business wants ready-to-post promotional content. A specialized model endpoint helps developers manage high-volume operational queues when they already maintain backend code and fallback rules. By contrast, a local shop, café, or retailer rarely wants custom software pipelines; it needs finished images, short videos, and captions built from existing phone photos without adding editing software or complex prompts to the workday.

About the author
Lan He
I'm Lan. Years of writing promotional and editorial work in the US, mostly for companies with big audiences and bigger expectations — where a single article might reach hundreds of thousands of readers, and where the good ones kept earning traffic long after the campaign ended. I learned what makes a page survive that: one clear audience, one real question, images that show the actual work, and a structure that keeps people reading. Now I write for small business owners who do everything themselves, answering the questions they actually ask: what's worth posting, which photos are worth taking, and how to tell one story across images, video, and words. I've seen good content carry a small company further than its ad budget ever did.
Give them a reasonto stop by.
Send us your phone photos. Get posts ready to share.
See what we can make
Keep exploring
All stories ↗- News
How to Use Inkling 2 Mini for a Product Photo Reshoot List
Give Inkling 2 Mini a product photo and one shopper’s question, then ask for a written list of views that would answer it. Hugging Face’s Transformers example takes a local image and a prompt together. A photo of a zipped pouch might show the pull clearly while leaving the inside a mystery. Check your existing photos before booking a reshoot. If you need a new shot, open the pouch and make sure its base is visible at posting size. Stillmore can use the checked photos and your product notes to make a finished social image or short video with a caption for you to publish.
- News
How to Use Gemini on Mac to Turn a Mascot Sketch Into a Blender Scene
To turn a mascot sketch into a Blender scene, choose Task in the Gemini app for Mac, mention @Computer use, and ask it to work from a copy of the drawing. Give the scene a filename and ask Gemini to stop after the first pose. Google began rolling out Computer use on November 19, 2026, to eligible adult subscribers using personal accounts in English in the US. Reopen the draft in Blender and compare the mascot's outline and markings with the drawing. Once you approve a pose, you can give Stillmore a reference frame and your product photos. Its team can make the promotional image or video and caption for you to publish.
- News
How to Use Gemini 4 Neon to Log Product Footage and Find Missing Shots
Ask Gemini 4 Neon to log each visible action, match the rows to your required shots, then replay the matches to decide what you still need to film. Supply the original clips, the approved product specifications and a brief that lists each required view. A faint click in a travel mug clip tells your editor where to look. If a thumb hides the latch, check another take before booking a side close-up. For a launch post built from product photos, Stillmore can plan, produce and check the visuals and supply a caption for you to publish.