Microsoft Decision-1 Scores AI Choices at $0.042 per Million Tokens

Microsoft has launched Decision-1, a model built to score, rank and select decisions rather than generate open-ended text. Announced on October 9, the model arrives as AI agents take over more enterprise work, and the quiet economics of agent infrastructure start to matter more than headline benchmark scores.

Unlike a general-purpose large language model, Decision-1 takes a state description and a fixed set of answer options, then returns a calibrated probability for each option. The design targets software that needs to act immediately on the result: routing a ticket, classifying a document, verifying a step in an agent workflow, or deciding which tool an agent should call next. According to coverage by TechDefused, Microsoft positioned the model directly against TypeSafe AI's Jev, the startup offering that helped define the emerging "decision model" category.

A model that scores instead of writes

The current release is post-trained from Qwen3.5-9B, the open-weight model from Alibaba Cloud, and is available in public preview through Microsoft Foundry and via OpenRouter. Microsoft says it plans to rebase future versions on its own MAI models, OpenAI systems, or other technology, according to reporting by MarkTechPost. The weights are not released; Decision-1 is offered only as a hosted API, with a 32,768-token context window and JSON output.

Five design goals guided the build, according to technical coverage of the release: low latency (around 85 milliseconds at the median in Microsoft's tests), generalization across held-out benchmarks, robustness to paraphrasing and reordering of the answer options, well-calibrated probabilities, and safety filtering that refuses harmful requests while preserving utility. The supported task shapes include yes/no questions, multiple-choice selections, ratings, and rubric-based grading of AI responses or agent actions.

Pricing is the part most likely to reshape agent deployments. Microsoft charges $0.042 per million input tokens, and output tokens are free. The company claims Decision-1 costs more than 20 times less than OpenAI's GPT-6 Sol for text-classification tasks. Because the model's answer is a choice rather than paragraphs of generated text, the most expensive part of a typical model call largely disappears.

The decision-model race heats up

The launch follows the path that Jev opened in mid-September, and Microsoft is not alone in chasing it. According to analysis by Dataist, OpenAI and Cloudflare have released similar offerings since Jev drew attention to the category, and Cloudflare's open Clef models are also built on Qwen. Decision-1 sits between those camps: hosted like Jev, built on Chinese open weights like Clef.

Microsoft reported 83.5 percent accuracy across 36 benchmarks covering roughly 150,000 questions, and said Decision-1 runs about 2.8 times faster than Jev. The numbers invite a caveat that Dataist flagged directly: Cloudflare's Clef models were not included in the comparison. The speed claim is therefore specific to the models Microsoft tested, not a measure against every relevant alternative. Every benchmark is vendor-run, which means buyers should treat the figures as a starting hypothesis for their own evaluation rather than a settled result.

The Qwen foundation also carries its own awkwardness. Microsoft building a flagship decision model on an open-weight model from China's Alibaba Cloud lands oddly for a frontier US technology firm, according to TechDefused, though Microsoft has been clear that the foundation is expected to change in later versions.

Why agent economics now hinge on small decisions

The strategic logic behind Decision-1 is straightforward. Most enterprise agent workflows are not a single brilliant answer. They are long chains of small decisions: route this request, classify that document, decide whether to escalate, verify that the last step worked. When every one of those steps runs through a large general model, the bill and the delay compound quickly.

A cheap, fast model dedicated to choosing can take over much of that traffic and leave the large models for the work that actually needs generation. Achint Srivastava, Microsoft's VP of Software Engineering in the Office of the CTO, described decision models as purpose-built to deliver structured outputs that software can act on immediately, in contrast to large language models designed to generate text or reason through complex problems. That distinction captures the bet: as agents multiply inside enterprises, the winners of the infrastructure layer may be the models that choose fastest and cheapest, not the ones that write best.

Read the TechDefused breakdown of Decision-1, the MarkTechPost technical overview, and the Dataist analysis of the benchmark gaps for more detail.