Microsoft Decision-1: Fast Agent Decision Model in Foundry
On October 9, 2026, Microsoft announced Microsoft Decision-1: a 9-billion-parameter model purpose-built to score fixed choices. It is now live in Microsoft Foundry at $0.042 per million input tokens, with output tokens free. The launch adds a major platform vendor to a category that barely existed six months ago: decision models.
This is not an LLM update. Decision-1 does not write, reason through open problems, or generate code. It takes a situation and a bounded list of answers, then returns a calibrated probability for each option in a single forward pass. That constraint is exactly what makes it fast, cheap, and suitable for the high-frequency decisions that enterprise AI agents make hundreds of times per workflow.
Why a Separate Model Category
LLMs are powerful generalists. They draft, reason, and synthesize. But using a frontier model to decide whether a support ticket should route to Tier 1 or Tier 2 is like hiring a senior engineer to sort the mail. The model is overqualified, slow, and expensive for that task, and calling it repeatedly across every agent step adds up fast.
Decision models invert the assumption. They do not try to understand language broadly; they score fixed options precisely. Give Decision-1 a situation and a list of four possible next actions, and it returns a probability distribution. The agent takes the highest-scored action, or sends the case to a human when nothing scores high enough.
The result is a check between what an agent knows and what it does, available as a cheap API call.
What Microsoft Decision-1 Does
Microsoft describes the model as designed for five use cases:
| Use case | What it does |
|---|---|
| Routing | Assigns incoming items (tickets, emails, requests) to the right queue, agent, or human |
| Classification | Labels content with a category from a fixed taxonomy |
| Prioritization | Scores items by urgency or importance |
| Verification | Checks whether an agent’s proposed output meets criteria |
| Workflow control | Decides the next step in a multi-step agent pipeline |
The interface is simple. Send a context block (up to 32,768 input tokens) plus a list of allowed options. Decision-1 returns a JSON object with a probability for each. Text in, JSON out, no free-form generation.
Achint Srivastava, VP of Software Engineering in Microsoft’s Office of the CTO, led the launch. He previously co-founded Pi Labs, an AI evaluation startup acquired by Microsoft, which built tools to assess and improve AI outputs. Decision-1 applies a related concern in a production model: scoring AI output against explicit criteria at inference time, as described in his official launch post.
Performance and Pricing
Microsoft benchmarked Decision-1 against 36 evaluations covering nearly 150,000 questions, all kept blind from training. It claims the highest accuracy in the comparison and the lowest latency:
- 4.5x faster than Quyet-1.0-Large, the previous fastest in the comparison
- 35x faster than GPT-6 Sol on structured decision tasks
- $0.042 per million input tokens, with output tokens free
To put the pricing in context: routing 1 million agent decisions costs $42, assuming short context windows. A single decision filling the full 32,768-token window costs less than two-tenths of a cent.
Inside Microsoft, two teams are already using Decision-1 in production. Xbox Research used it to classify more than 10,000 feedback items. The Copilot team used it to assess AI response quality. Both are classic enterprise decision tasks: scoring something against a fixed rubric at volume, where latency and cost matter as much as accuracy.
The Broader Decision Model Category
Decision-1 is not alone. TypeSafe AI raised $870 million after its non-text model Jev went viral in September, with a third of Fortune 500 companies reportedly using it within weeks of launch. Cloudflare’s Clef offers open-weight decision scoring via Workers AI. Perplexity and Liquid AI have introduced models in the same category.
The pattern is consistent across all of them: fixed inputs, structured scored output, no free-form generation. TypeSafe positions Jev as a language for automation rather than communication. Microsoft positions Decision-1 as the check between an agent’s reasoning and its next action.
The practical difference for enterprise buyers is ecosystem fit. Decision-1 lives in Microsoft Foundry, which means organizations already running models through Foundry can add decision scoring to their pipelines without a new vendor relationship. Clef is open-weight and available without a cloud dependency, but teams that rely on Foundry’s cost controls and governance tooling will likely test Decision-1 first.
OpenRouter access is planned but not yet dated, which means adopting Decision-1 today ties your escalation thresholds to Microsoft’s endpoint until that arrives. Moving later means recalibrating against your own labeled data.
What Enterprise AI Builders Should Know
Decision models are additive, not replacements. Decision-1 belongs in the same pipeline as an LLM, not instead of one. The LLM reasons and generates; the decision model scores and routes. The combination handles both open-ended and bounded tasks without overloading the expensive model.
You own the evaluation set. The real adoption cost of Decision-1 is not the $0.042 per million tokens; it is the labeled data that defines what “correct” means for your specific decisions. Microsoft’s benchmarks show strong general performance, but your routing rules and escalation thresholds are specific to your context. Budget for calibration time before going to production.
Human oversight stays your responsibility. Microsoft’s documentation is explicit: Decision-1 should not decide alone on employment, credit, housing, healthcare, or legal-rights cases. The model scores options; your team defines the thresholds and maintains the human review loop for high-stakes calls.
Context window is 32,768 tokens. That is sufficient for most routing and classification tasks, but not for long document review or multi-turn conversation assessment. Knowing the limit before designing a pipeline avoids costly redesigns.
For teams already managing agent governance with a unified control plane, a decision model is a natural addition: one scored call per decision point, logged and auditable, with thresholds you control. For teams building durable agent workflows, routing and verification calls are exactly the kind of low-latency, high-frequency checks that a 9B model handles better than a 405B generalist.
What This Signals
The commercial AI market is splitting by task type. Frontier models get more powerful and handle open-ended reasoning. A new category of small, fast, structured models handles decisions, routing, and verification at scale. Both will be in production enterprise stacks simultaneously, doing different jobs.
Microsoft entering the decision model category with a Foundry-native product signals that the category is now a product line, not an experiment. The benchmarking is Microsoft’s own, and independent validation will come over the following months. But the pricing, the platform integration, and the early internal deployments all point to a product ready for serious evaluation.
If your agents make dozens of routing, classification, or verification calls per workflow, running those calls through Decision-1 instead of a frontier model is the simplest cost-reduction and latency-reduction move available today. The risk is calibration: test your thresholds against your own data before trusting Microsoft’s general benchmarks in production.
For more on how enterprise AI infrastructure is maturing, see Claude in Azure Foundry and the agentic enterprise stack.