Microsoft Decision-1 is Microsoft's new model for fast, structured choices inside software. On October 9, 2026, Satya Nadella and Microsoft engineering leads announced the model in Microsoft Foundry, with OpenRouter access planned soon. For India and USA agencies shipping agent workflows, this is a control-layer story, not another chatbot launch.

Microsoft Decision-1 facts: job, speed, price, Foundry, stability — Microsoft Decision-1 claims at a glance. Graphic: Debugged Pro original. Facts from Microsoft Foundry coverage and RuntimeWire.

Clients will ask whether every support ticket, agent tool call, or quality check should now hit a decision model. Your job is to map the product claims to one reversible pilot, with logs and a human path.

What Microsoft announced about Microsoft Decision-1

According to The Verge, Microsoft launched Microsoft-Decision-1 just days after OpenAI highlighted its Decisions API. The point of these models is to choose something quickly and efficiently, such as which model to call next or how to label an input.

RuntimeWire reports that Microsoft VP Achint Srivastava framed Decision-1 for routing, classification, prioritization, verification, and workflow control. Given a situation and a defined set of choices, the model returns a probability for each option.

Base model: post-trained from Alibaba's open-weight Qwen3.5-9B (Microsoft-reported)
Availability: Microsoft Foundry now; OpenRouter coming soon
Pricing: $0.042 per million input tokens; output tokens free
Context: text inputs up to 32,768 tokens; JSON returns
Not supported: open-ended generation, images, audio, or video

Microsoft says future versions may rebase onto Microsoft MAI and OpenAI models. Treat that as a product roadmap note, not a shipping guarantee for your current sprint.

Why Microsoft Decision-1 matters for agency delivery

Agencies do not need another model card. You need fewer brittle JSON parses and clearer gates between an agent's plan and its next tool call.

Three pressures show up in India and USA accounts this week:

Client FOMO. Product owners saw Nadella's post and want "decision models" on the roadmap.
Cost pressure. If a check costs fractions of a frontier chat call, finance will ask why every verification still burns GPT-class tokens.
Latency stacking. Microsoft notes that adding 100 milliseconds to each of 20 dependent decisions adds two seconds to a workflow. Serial agent loops feel that delay.

This sits next to production trust questions we covered on AI agent production trust and review load in AI coding agents and review bottlenecks. Faster choices still need process changes.

Benchmarks Microsoft published (vendor-reported)

Microsoft says Decision-1 led a comparison across 36 blind benchmarks covering nearly 150,000 questions withheld from training. In those tests, the company reports:

Highest accuracy among the models it evaluated
About 35 times faster median latency than GPT-6 Sol
About 4.5 times faster than runner-up Quyet-1.0-Large
Decision flips in only 1.3% of cases across eight prompt perturbations; no flips when option order was reversed or shuffled
Safety testing across 5,250 requests on 11 benchmarks covering harmful content, jailbreaks, and prompt injection

These numbers come from Microsoft. Independent reproduction is limited so far. Put them in decks only with that caveat, then replace them with your labeled sample results.

Internal Microsoft trials worth noting

Xbox Research: labeled more than 10,000 feedback items; Microsoft says quality was competitive with GPT-6 Sol while running over 14 times faster and about 200 times cheaper.
Copilot team: used the model for quality checks; Microsoft says it was competitive with GPT-5.6 Luna and about 100 times faster.
Other uses: Microsoft cites incident knowledge retrieval and scientific replanning, with scoring said to be more consistent than an LLM-based approach.

How Microsoft Decision-1 fits next to TypeSafe Jev and OpenAI Decisions

Decision models are becoming a category, not a one-company stunt. Cloudflare shipped Clef models earlier in October. OpenAI's Decisions API is in public beta. TypeSafe's Jev drew heavy attention after its September launch and October funding news, which we covered in our TypeSafe AI Jev guide.

For agency architects, the useful split is:

Chat / reasoning models: plan, explain, draft, explore
Decision models: pick among fixed options with a confidence score
Classic classifiers: still fine when labels are stable and volume is high

Do not force every branch through a long chat prompt. Do not throw away a working rules engine for a press release either. Pick the thin layer that matches the decision contract.

Agency checklist for Microsoft Decision-1 pilots — Six-step agency checklist before production use of Microsoft Decision-1. Graphic: Debugged Pro original.

A practical Microsoft Decision-1 pilot for India USA agencies

1. Write the decision contract

List allowed labels, confidence thresholds, and what happens on "cannot tell." If you cannot fit that on one page, you are not ready to automate the branch.

2. Choose one reversible flow

Good first fits: support ticket routing, lead band scoring, content flags before a human queue, or an agent tool-gate that only allows safe tools. Poor first fits: employment, credit, housing, healthcare, or legal rights. Microsoft's own Foundry notes warn against sole automated use for those consequential paths.

3. Shadow mode for two weeks

Log the model's choice beside your current system. Do not let it act yet. Hold a weekly review with one engineer and one client stakeholder.

4. Calibrate on your data

Pull 500 to 2,000 labeled examples from the last quarter. Score Microsoft Decision-1 against your current parser or rules. Measure accuracy, abstention rate, and p95 latency from India and U.S. regions if you serve both.

5. Price the full path

Include input tokens, retries, human review minutes, and incident cost. A cheap model that needs heavy cleanup is not cheap.

6. Keep NIST-aligned language in proposals

Map Govern, Map, Measure, and Manage from the NIST AI Risk Management Framework. U.S. buyers already expect that framing when AI sits in production paths.

Where Microsoft Decision-1 helps client roadmaps

If your studio sells software development consulting or AI automation services, add a decision-layer audit to discovery. Ask:

Which steps today parse chat text into enums?
Which agent actions lack a confidence gate?
Which queues burn senior review time on easy labels?
Which flows need multi-region latency checks?

India delivery teams serving U.S. clients should document data paths into Foundry and any residency constraints. USA product agencies should ask procurement about audit logs, model version pinning, and fallback when Foundry or OpenRouter is slow.

Risks and caveats to say out loud

Benchmarks are Microsoft-reported. Ask for datasets and baselines before you put 35x claims in a client deck.
Narrow scope by design. No explanations, no multimodal inputs. Pair with a reasoning model when users need a why.
Calibration is your job. A probability only helps if your product knows when to abstain or escalate.
Category is crowded. OpenAI Decisions, Cloudflare Clef, and TypeSafe Jev will keep price and feature pressure high. Revisit contracts quarterly.
High-stakes bans still apply. Do not be the sole automated decider for rights-affecting outcomes.

Clients respect agencies that refuse hype. You can still run a serious Foundry pilot without promising that Microsoft Decision-1 replaces your whole stack.

Suggested 30-day rollout

Week 1: Pick one flow. Write the decision contract. Pull labeled history. Confirm Foundry access and billing alerts.

Week 2: Integrate behind a feature flag. Log input hash, choice, confidence, latency, and model version.

Week 3: Shadow mode only. Compare against current rules or chat parsing. Hold a human review board.

Week 4: If accuracy and p95 clear your bar, open to 10% of traffic with instant rollback. Document results for the client.

Budget one senior engineer and one QA owner. Do not staff this as a weekend hack if the client is enterprise.

Key Takeaways

Microsoft Decision-1 launched October 9, 2026 in Microsoft Foundry as a structured decision model, with OpenRouter planned.
It returns calibrated probabilities for fixed choices at $0.042 per million input tokens, with free output tokens.
Microsoft reports large speed and accuracy gains versus GPT-6 Sol and other decision models; treat those as vendor results until you reproduce them.
India and USA agencies should pilot low-risk routing or agent gates with shadow mode, labeled tests, and human overrides.
Decision models sit beside chat and coding agents. Use them for branches, not for open-ended writing.

FAQ

What is Microsoft Decision-1?

Microsoft Decision-1 is a purpose-built model that scores fixed choices instead of generating open-ended text. Given a situation and defined options, it returns a calibrated probability for each choice through a structured API. It is available in Microsoft Foundry, with OpenRouter access planned.

How much does Microsoft Decision-1 cost?

Microsoft lists input tokens at $0.042 per million, with output tokens free. That pricing positions Decision-1 as a low-cost control layer for repeated routing, verification, and agent action checks inside production workflows.

How is Microsoft Decision-1 different from a chat LLM?

Chat LLMs generate text. Microsoft Decision-1 is built for yes/no, multiple-choice, ratings, and rubric grading. It returns structured scores software can act on without parsing paragraphs. Microsoft says it does not generate explanations or accept images, audio, or video.

Should India and USA agencies adopt Microsoft Decision-1 now?

Treat it as a strong pilot candidate for reversible flows such as ticket routing or agent tool gates. Calibrate on your labeled data, keep human review for high-stakes paths, and note that published speed and accuracy numbers are Microsoft-reported until you reproduce them.

Where can agencies try Microsoft Decision-1?

It is available now through Microsoft Foundry. Microsoft says OpenRouter access is coming soon, without a public date. Teams already on Azure can start with a feature-flagged shadow deployment behind one client workflow.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA agency teams on AI delivery, software process, and client-ready rollouts. More from the team on the Debugged Pro blog and about us page.