By Amara Osei

Microsoft has launched Decision-1, an AI model that returns calibrated probabilities for a fixed set of answers instead of generating free-form text — and it prices that capability at 4.2 cents per million input tokens, with no charge for output tokens.

The model is available now through Microsoft Foundry and OpenRouter. Microsoft built it by post-training Qwen3.5-9B and says it plans to rebase it on other models, including Microsoft AI (MAI) and OpenAI models.

The design targets a specific gap in enterprise AI stacks: the routine decisions inside agent workflows — classifying a request, ranking incidents, approving a step — that today often run through a general-purpose large language model. Those calls produce text that software must then parse. Decision-1 skips that step.

"Instead of writing a free-form response, it takes a fixed set of answer options and returns a calibrated probability for each one in a single pass," Microsoft said in a blog post.

The mechanism is simple to state. Feed the model a question and a set of possible answers. It returns a probability for each option. It can also evaluate multiple decisions in parallel within one request. Developers can wire those outputs directly into code to classify inputs, rank work, verify AI-generated responses, or pick an agent's next action.

In short: this is a model for verdicts, not prose.

What do the benchmarks and the price actually mean?

Microsoft claims Decision-1 achieved the highest accuracy across 36 benchmarks covering nearly 150,000 questions kept blind from training. The company also reported the model ran about 2.5 times faster than the runner-up and 35 times faster than GPT-6 Sol.

The cost frame matters more for buyers. In an incident-prioritization example, Microsoft said a request using 86 input tokens and two output tokens cost roughly $0.0000036. At that rate, one dollar covers more than 275,000 similar decisions.

To translate: routing and triage calls that would otherwise hit a generative model's token bill can drop to fractions of a cent each. For an organization running millions of agent interactions per year, that arithmetic compounds quickly.

One caveat: the benchmark results and cost calculations are Microsoft's own claims and have not been independently verified.

Why are analysts cautious?

Analysts see a real architectural shift, with strings attached.

Charlie Dai, vice president and principal analyst at Forrester, said specialized decision models could reduce the cost and latency of agentic AI workflows. "Routing, classification, and approval checks often represent a large share of agent interactions, so a specialized decision model can significantly reduce cost and latency while reserving general-purpose LLMs for reasoning and content generation," Dai said.

Amit Kumar Jena, development manager for AI at Kanerika, called the move of routine routing and approval checks away from large generative models a meaningful architectural change for enterprises scaling AI agents. But he cautioned against using Microsoft's benchmark numbers to forecast enterprise savings. Organizations should measure "cost and latency per successfully completed workflow, including fallbacks and mistakes," he said.

What are the governance risks?

Decision-1 returns scores without a rationale. Microsoft's documentation also warns that results can change depending on how questions are worded, Jena noted.

"My concern is that a precise-looking probability can give an uncertain decision more authority than it deserves," he said.

Dai said the absence of reasoning makes decisions harder to explain, validate, and challenge. He recommended:

• calibration testing before deployment
• controls over question wording and score thresholds
• audit logging of decisions
• human review for high-impact decisions, especially where customer outcomes or regulatory obligations are involved

Jena advised starting with reversible tasks such as request routing and triage, and requiring meaningful human review for any decision that materially affects people.

The core issue sits with interpretation, not the model itself. A probability assigned to an answer does not explain the decision or prove it is correct for a specific enterprise application.

For CIOs, Decision-1 is a cost lever — one that demands governance before it touches anything consequential.