Microsoft's Decision-1 Is a Post-Trained Qwen3.5-9B at $0.042 Per Million Tokens, With Self-Reported Benchmarks

Microsoft released Microsoft-Decision-1 on October 9, a model post-trained from Alibaba's Qwen3.5-9B that returns a probability for each of a fixed set of answers instead of generating text. It costs $0.042 per million input tokens, and output tokens are free.

Achint Srivastava, a vice president of software engineering in Microsoft's Office of the CTO, announced it in a post on Microsoft Command Line. It is available in Microsoft Foundry and on OpenRouter. Microsoft's post describes access through those two services only and does not say whether the weights are downloadable.

What Decision-1 does

The model handles yes-or-no questions, multiple choice and ratings, and it can grade AI responses or agent actions against a rubric. Microsoft lists routing, classification, prioritization, verification and workflow control as uses. It takes text only, with a 32,000-token context window according to The Next Web, and it gives no explanation with its score.

Srivastava wrote that "Decision models are quickly emerging as an important new category in AI." Microsoft also said it plans to rebase the model on other models, including its own MAI models and OpenAI's.

The numbers are Microsoft's

Microsoft says Decision-1 had the highest accuracy in a comparison across 36 benchmarks covering nearly 150,000 questions kept blind from training. It says the model is 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1, and 35 times faster than OpenAI's GPT-6 Sol.

The Next Web notes that the 35 times figure compares a scoring model with a text generator. None of the coverage reviewed includes an independent reproduction.

Claim (all Microsoft's)	Figure
Input price per million tokens	$0.042
Decision changes under rewording or reordering	1.3% on average
Safety test set	5,250 requests, 11 benchmarks
Xbox Research feedback sorting vs GPT-6 Sol	14x faster, 200x cheaper

The Xbox Research result covered more than 10,000 pieces of player feedback. Microsoft said the model was competitive with GPT-6 Sol on quality there, and the Copilot team found it competitive with GPT-5.6 Luna at 100 times the speed.

What analysts flagged

InfoWorld reported on October 13 that Charlie Dai, a vice president and principal analyst at Forrester, said specialized decision models could cut the cost and latency of agent workflows. He recommended calibration testing, audit logging and human review for high-impact decisions.

Amit Kumar Jena, a development manager for AI at Kanerika, cautioned that "a precise-looking probability can give an uncertain decision more authority than it deserves." Microsoft's model card says the model should not be the sole basis for consequential decisions about people, such as credit or employment.

The field is crowded. OpenAI's Decisions API is in public beta, and The Next Web, citing The Register, reports that Cloudflare, Strands, Liquid AI, Perplexity and Snowflake have announced decision models of their own. Microsoft later updated its post with benchmarks against Jev from TypeSafe AI.

Small open models are already being reshaped this way elsewhere, as in WorldTUI's 360M-parameter terminal simulator, and cheap routing models sit under the price cuts covered in DeepSeek's V4.1-Flash release.

Microsoft has not published a date for the rebase onto MAI or OpenAI models, and the price may change with it, since the company describes pricing as dynamic.