Microsoft-Decision-1 bills input only, at $0.042 per million
Microsoft on October 9 made Microsoft-Decision-1 available in Microsoft Foundry. The model scores a fixed list of answers rather than writing a reply, and it bills $0.042 per million input tokens with output free, so for teams routing agent steps or labelling data the price of each decision depends only on the prompt: the instructions, the item being judged and the options.
In Microsoft’s announcement, the company says it post-trained Qwen3.5-9B for single-pass scoring and will soon rebase the model on others, including Microsoft AI (MAI) and OpenAI models. You pass in a set of options (yes/no, multiple choice or a rating) and get back a calibrated probability for each one. The same structured API call also handles rubric-based grading of AI responses and agent actions.
That probability is part of the API. Microsoft expects applications to use it to decide when to act, when to defer and when to ask for human review, and its stated bar is that a 90% prediction should be right about nine times out of ten on representative cases. Latency is the other half of the pitch, and Microsoft gives an example of how it compounds: adding 100 milliseconds to each of 20 sequential decisions adds two seconds to the workflow.
Microsoft suggests a long list of uses. Among them: whether an agent should continue, stop, retry or hand off; which model should take an incoming request, based on quality, cost and latency; whether to accept, revise or reject an AI-generated answer against set criteria; and how to apply an agent skill’s rules without reprocessing a long list of instructions every time.
Who it was measured against
Microsoft took several of the top models from JevBench, a public leaderboard, and tested them on 36 public and private benchmarks with nearly 150,000 questions, all kept out of training. Microsoft-Decision-1 scored the highest accuracy. It was also the fastest model measured, 4.5 times quicker than the runner-up, Quyet-1.0-Large, and about 35 times quicker than GPT-6 Sol at median (P50) latency. The ranking comes only from Microsoft’s own tests. The models Microsoft lists as benchmarked include OpenAI’s GPT-6 Luna Decisions and AWS’s Strands-Decider 2B, along with Quyet-1.0-Large, H2O-Lightning-4B, deck-31B and Surogate Rune 26B-A4B.
Robustness was tested separately. Microsoft rewrites each request in eight ways, paraphrasing instructions, rewording or reordering options, changing keys and adding formatting noise, then counts how often the decision flips. The average is 1.3%, with zero flips when option descriptions are paraphrased or when options are reversed or shuffled. On safety, Microsoft ran 5,250 requests from 11 benchmarks covering harmful content, jailbreak attempts and prompt injection. It says the model refused harmful behavior while keeping a high degree of utility.
Strands-Decider 2B, also on that list, runs locally. AWS took Qwen3.5-2B, replaced the head that writes text with a small pointer head that scores the options a developer supplies, and released it on October 1 as a download along with its training data and scripts. AWS says it decides in under 100 milliseconds on an Nvidia RTX 3090. Microsoft-Decision-1 is offered through Foundry.
Trials inside Microsoft
Xbox Research ran more than 10,000 open-ended pieces of feedback and reviews from surveys, Steam and Twitter/X through the model, sorting them into themes its researchers had set. The team found it competitive with GPT-6 Sol on quality while running over 14 times faster at 200 times lower cost.
The Copilot team, which measures the quality of chat and agentic responses, found it competitive with GPT5.6 Luna and 100 times faster.
Microsoft Discovery’s agents grade a previous experiment against a rubric, revise the approach and repeat until the objectives are met. There, Microsoft-Decision-1 scored 46 times more consistently than the LLM-based grader at three times the speed, and made adaptive replanning nearly four times faster.
To try the model, open its Foundry catalog page. Microsoft says OpenRouter access is coming soon.
Up next
USA Today wants over $250 million from OpenAI
“Real and continuing” harm is how USA Today describes OpenAI training on “hundreds of thousands” of its articles, and it wants damages exceeding $250 million in Manhattan federal court.
Related stories
- Perplexity puts its 4B decision model on the Mac
- Microsoft AI’s MAI-Base-2 trained on 16K GPUs with Muon
- SubQ Snap reads 384K tokens before each agent decision
- OpenAI’s Decisions API beats its own engine by up to 10x
- AWS’s Strands Decider 2B picks from your list, never writes
- Linear and Ramp users can pay for Claude without an API key
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

