Microsoft-Decision-1 is a small model built for the boring decisions inside an agent loop: announced on Microsoft’s Command Line blog on October 9, 2026 and listed on Product Hunt on October 10, 2026, it takes a fixed set of options and returns a calibrated probability for each one in a single pass, as structured output that code can act on directly. Microsoft says it is post-trained from Qwen3.5-9B, about 35x faster than GPT-6 Sol at median latency, and priced at $0.042 per million input tokens in Microsoft Foundry with output free.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Microsoft-Decision-1: A Fast, Cheap Model for Agent Routing Calls |
| Short answer | Microsoft-Decision-1, announced on the Command Line blog on October 9, 2026, is a decision-scoring model for routing, classification, verification and agent control. |
| Reading time | ~3 min |
| Last updated | 2026-10-12 |
What it ships with
- Decision scoring, not chat. Give it fixed options and it returns a calibrated probability per option, which is what routing, classification and verification actually need.
- Structured output. The result is machine-readable, so an agent can branch on it without parsing prose.
- Speed. Microsoft’s figure is roughly 35x faster than GPT-6 Sol at median latency.
- Price. $0.042 per million input tokens, free output, available in Microsoft Foundry.
- Base model. Post-trained from Qwen3.5-9B, according to the listing.
- Launch numbers. Day rank 4, 144 points and 118 followers at the time of writing.
Why the small decisions are the expensive ones
A desktop agent does not spend most of its budget on the hard step, it spends it on the dozens of easy steps around it: is this the file the user meant, did the command succeed, should this be retried, is this tool call safe. Answering those with a frontier model is like hiring a specialist to read a checklist. A narrow model that returns a number per option, quickly and cheaply, is the more sensible shape for that job.
There is a real caveat. A calibrated probability is only as good as the options you hand it, and a model that scores decisions is not a model that understands your machine. If the wrong file is among the candidates, a confident probability does not save you.
What this means for a Windows agent
Fast, cheap decision scoring is the part of the stack that makes an approval-first desktop agent practical, because the agent can verify its own small steps instead of asking you about every one. The execution side still has to happen on your PC, which is where a Windows-native agent comes in: MOOGH proposes a plan, waits for your approval and then works on your own files and commands, and it is a free download.
What the official announcement adds
Microsoft’s own post is where the numbers come from, and it fills in three things a launch listing blurs.
- Who and when. Achint Srivastava, VP of Software Engineering in Microsoft’s Office of the CTO, published the announcement on the Command Line blog on October 9, 2026.
- Where you can call it. Microsoft Foundry and OpenRouter, both with documentation.
- How the score is produced. The model handles yes/no, multiple-choice and rating options as well as rubric-based scoring, and returns a calibrated probability per option — a single-pass scorer rather than a chat model.
The benchmark claim is narrower than the headline suggests: Decision-1 took the top accuracy slot in Microsoft’s own 36-benchmark comparison, spanning roughly 150,000 questions kept blind from training, and it was the fastest model Microsoft measured — 2.5 times quicker than H2O-Lightning-4B v1.1, the runner-up, and about 35 times quicker than GPT-6 Sol at median latency. Those are vendor-run benchmarks, which is a reason to test it on your own traffic rather than to dismiss it.
Pricing is the half that makes the pattern obvious: $0.042 per million input tokens, with output free. When a decision is a routing call between four options, paying frontier-model rates for it was always the awkward part of building an agent.
What to watch next
- Whether the speed claim holds outside Microsoft’s own benchmarks.
- How the latency and price compare with other small classifiers once third parties test it.
- Whether Foundry exposes the calibration numbers well enough to audit routing decisions.
FAQ
- Is this a replacement for Claude or GPT? No. It is aimed at narrow decision calls inside a workflow, not open-ended generation.
- Can I call it from my own tooling? It is listed as available in Microsoft Foundry, so access follows Foundry’s terms and pricing.
- Do I need it to build a desktop agent? No, but a cheap verification model makes approval-first loops much less annoying to run.
Related reading: