Models
What Is Microsoft Decision-1? The AI Model That Only Picks Answers
Microsoft Decision-1 is a decision model built on Qwen3.5-9B that returns a probability for each answer instead of text. Price, speed claims and Jev compared.
Microsoft Decision-1 (also written Microsoft-Decision-1 or Decision 1) is a new kind of AI model from Microsoft that does not write sentences. You give it a question and a short list of possible answers. It gives back a number for each answer: how likely that answer is to be the right one.
That sounds small. It is aimed at one of the biggest hidden costs inside AI agents: the thousands of tiny yes-or-no calls they make while they work. Here is what Decision-1 is, what it costs, how it compares with Jev, and what to make of Microsoft’s speed claims.
What is Microsoft Decision-1?
Most AI models you know, like ChatGPT or Claude, are language models. They produce words. Decision-1 produces a choice. Microsoft describes it as a model that takes "a fixed set of answer options and returns a calibrated probability for each one in a single pass."
It handles four kinds of questions:
- Yes or no. "Is this email spam?"
- Pick one. "Which team should get this support ticket: billing, shipping or tech?"
- Ratings. "How urgent is this incident, from 1 to 5?"
- Rubric grading. "Did this AI answer follow the rules?" or "Was that agent action correct?"
"Calibrated" is the key word. Microsoft’s own explanation is that a 90% prediction should be right about nine times out of 10. If that holds, software can set a simple rule, such as "act alone above 95%, ask a human below it."
The announcement came from Achint Srivastava, a vice president of software engineering in Microsoft’s Office of the CTO, who wrote that decision models "are quickly emerging as an important new category in AI."
Microsoft Decision-1 at a glance
| Question | Answer |
|---|---|
| Who made it? | Microsoft |
| Announced | October 9, 2026 |
| Built on | Alibaba’s Qwen3.5-9B, post-trained by Microsoft |
| What it returns | A probability for each answer option, as structured output |
| Context window | About 32,000 tokens, text only |
| Price | $0.042 per million input tokens, output tokens free |
| Where to use it | Microsoft Foundry and OpenRouter (hosted, for developers) |
| Open weights? | No download offered; it is a hosted model |
Microsoft also said it plans to rebuild Decision-1 on other base models, including its own MAI models and OpenAI models. It did not say why or when.
How much does Microsoft Decision-1 cost?
The price is the headline. Decision-1 charges 4.2 cents per million input tokens, and the answers it sends back are free. Tokens are the small chunks of text a model reads; a short support ticket might be under 100 of them.
InfoWorld walked through one of Microsoft’s own examples: an incident-priority check with 86 input tokens and 2 output tokens costs about $0.0000036. At that rate, $1 covers more than 275,000 similar decisions. For comparison, the cheapest tier of OpenAI’s GPT-5.6 family, Luna, lists $1 per million input tokens, about 24 times more, before you pay anything for output.
So is Microsoft Decision-1 free? No, but for a single person it rounds to almost nothing. It is built for companies making millions of these calls.
Microsoft’s speed and accuracy claims
Microsoft published a lot of numbers. All of them come from its own testing:
- Accuracy: the highest score in a comparison across 36 benchmarks with nearly 150,000 questions held out of training.
- Speed: about 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1, and about 35 times faster than GPT-6 Sol.
- Consistency: it changed its answer on 1.3% of requests, on average, when the same question was reworded or the options were reordered.
- Safety: tested on 5,250 requests across 11 safety benchmarks, including jailbreaks and prompt injection.
Microsoft also shared internal results. Xbox Research used it to sort more than 10,000 pieces of player feedback into themes, with quality Microsoft called competitive with GPT-6 Sol at over 14 times the speed and about 200 times lower cost. The Copilot team used it to check chat and agent answers, and found it competitive with GPT-5.6 Luna and about 100 times faster.
The honest caveat: nobody outside Microsoft has reproduced these figures yet. Forrester analyst Charlie Dai told InfoWorld that because Decision-1 gives a score with no reasoning, its choices are harder to explain, and he recommended audit logs and human review for high-impact decisions. Microsoft’s model card itself says it should not be the only basis for big decisions about people, like credit or hiring.
Microsoft Decision-1 vs Jev
Decision-1 did not invent this idea. Jev, from San Francisco startup TypeSafe AI, went into early access on September 15, 2026, and calls itself a "System One" model. Microsoft added Jev to its latency comparison after launch and said Decision-1 was 2.8 times faster, according to The Next Web, citing The Register.
| Microsoft Decision-1 | Jev | |
|---|---|---|
| Maker | Microsoft | TypeSafe AI |
| First available | October 9, 2026 | September 15, 2026 (early access) |
| Base model | Qwen3.5-9B, disclosed | Not disclosed |
| Returns | Probability per option | Choice, score or yes/no with confidence |
| Where | Microsoft Foundry, OpenRouter | TypeSafe’s own platform |
| Speed claims | Microsoft’s own tests | TypeSafe’s own tests |
The field is getting crowded fast. Inception says its Mercury Decide model, announced with Mercury 3, beats Jev on JevBench, and The Next Web reports that OpenAI has a Decisions API in public beta. When Microsoft and OpenAI both ship something, it usually stops being a niche.
Why decision models matter for AI agents
An AI agent works in a loop: plan, act, check, repeat. Most of that loop is deciding, not writing. Which tool comes next? Did that step work? Is this good enough, or should a person look?
Today each of those small calls often goes to a large model that writes a paragraph, which software then has to read. A decision model answers in one short step, for a fraction of the cost. Expect the agents you use to get faster and cheaper this way, with big models saved for the hard parts.
What this means for you
You will never open Decision-1. There is no chat window and no app. It is plumbing for developers, and it only does one thing: pick from a list.
What you actually want is the layer on top, where the deciding, routing and checking are already handled and you just get the finished work. That is the idea behind agentic AI. Agentic AI gives you a team of expert agents on your iPhone and iPad. Pick a specialist, or put up to 4 agents on one task, and they plan and finish it together. No account and no login, and the first 3 tasks are free.
Quick answers
What is Microsoft Decision-1? A Microsoft AI model, announced October 9, 2026, that scores a fixed list of answers and returns a probability for each one instead of writing text. It is used for routing, classification, ratings and checking agent actions.
Is Microsoft Decision-1 built on Qwen? Yes. Microsoft post-trained Alibaba’s open-weight Qwen3.5-9B. It says later versions will be rebuilt on other models, including its own MAI models and OpenAI models.
How do I use the Microsoft Decision-1 API? Through Microsoft Foundry or OpenRouter. You send the question and the answer options, and it returns structured output with a probability for each option. Documentation is on Microsoft Learn.
Microsoft Decision-1 vs Jev: which is better? Microsoft says Decision-1 was 2.8 times faster than Jev in its own test. Neither company’s figures have been checked independently, so test both on your own data.
Can I chat with Microsoft Decision-1? No. It does not produce sentences. If you want AI that writes, researches or plans for you, use an assistant or an agent app.
Sources: Microsoft: Microsoft-Decision-1 · Microsoft Community Hub · The Next Web · InfoWorld. Performance figures are Microsoft’s own and are not independently verified.
