“No Reasons, Just Probabilities”: How Microsoft-Decision-1 Swaps an LLM Out of the Agent “Checkpoint” for a 9B Judge
Microsoft has released Decision-1, a model built purely for judging. It writes no prose—it returns only calibrated per-option probabilities as JSON. Input costs $0.042 per 1M tokens; output is free. It lets you cheaply carve out the “checkpoints” of agent proceed/stop decisions, model routing, and A
If you've ever built a coding agent, you've inevitably run into this kind of work somewhere along the way. “Should this agent's next move be allowed to proceed, be stopped, or be handed off to a human?” “Which model should this request go to?” “Does this review result pass or fail?” Until now, these small “decisions” have been handled as a side job by expensive large language models (LLMs). Microsoft-Decision-1, released by Microsoft on October 9, 2026, is a small model that carves out just this decision-making role. It writes no prose—it simply returns a probability for each option you give it.
Stop Making It Write, Make It Return Probabilities
A typical LLM generates free-form text in response to a question. Decision-1 doesn't. When you hand it a predefined set of options (yes/no, multiple choice, a rating score, rubric-based grading, and so on), it computes a calibrated probability for each option, returns it as JSON, and stays silent. No reasoning, no excuses attached. The design lets your application look at those numbers and branch directly into “run automatically,” “hold the decision,” or “send it to a human reviewer.”
Under the hood it uses the 9B-parameter version of Alibaba's open model Qwen3.5, which Microsoft fine-tuned for one-shot scoring. The input is text, and it can reportedly accept up to 32,768 tokens. On benchmarks it ranked first across roughly 150,000 questions and 36 types of tests, and it also demonstrated stability: when small perturbations (such as changes in phrasing) are added to the input, the decision changes on average only 1.3% of the time.
A Price Tag on the “Orders of Magnitude Cheaper” Side
The biggest change this separation brings is cost and speed. The fees and latency that used to be incurred every time an LLM was called as a judge drop sharply once they're replaced by a dedicated model. The figures Microsoft published are as follows.
| Item | Microsoft-Decision-1 |
|---|---|
| Input pricing | $0.042 per 1M tokens |
| Output pricing | Free |
| Speed (P50 latency) | About 35× faster than GPT-6 Sol |
| Model size | 9B parameters (fine-tuned Qwen3.5-9B) |
Achint Srivastava of Microsoft (co-founder of Pi Labs), who led the development, positions this as a “low-cost control layer for agentic systems.” The idea is that confidence scores let software decide whether to “run, hold, retry, escalate, or hand off to a person or another tool.”
Where It Helps in a Coding Agent
In setups that run agents like Claude Code or Codex unattended, the decision model can be slotted in as a “checkpoint” like the following. The use cases are listed officially; pulling out the ones most relevant to development workflows, we get the following.
Next-Move Gate for Agents
Evaluate the next action an agent proposes and decide—by probability—whether to proceed, stop, retry, or hand off to a human. You can place a cheap “brake” just before a runaway process or infinite loop.
Model Routing
Look at an incoming request and pick the model that best fits the quality, cost, and latency requirements. You can do the triage up front—sending simple tasks to small, cheap models and only the hard ones to expensive models.
AI Judging and Validation (Replacing LLM-as-a-Judge)
Grade whether an output or review result passes or fails according to a rubric. The step where an LLM used to double as the “judge” can be moved to a dedicated grader.
In every case, the idea is to decouple the part that used to call a large model for each decision. Given the recently talked-about survey finding that “AI increased code generation, but the review stage got congested,” the significance of having more options to cheaply add on automated checking and validation is far from small.
What to Keep in Mind Behind the Convenience
At the same time, a design that reduces decisions to numbers carries its own concerns. Because Decision-1 returns no reasoning, its auditability—being able to trace after the fact “why it arrived at that probability”—tends to be weaker than an LLM that explains itself in prose. Leaving important branches to probability scores alone makes the basis for a decision prone to becoming a black box, which calls for caution. And “calibrated” or not, a probability is still a probability; depending on where you set the threshold, misses or excessive holds can occur. The more a checkpoint touches on safety, the wiser it is not to take scores at face value but to combine them with human review and log cross-checks. Making the judge role cheaper and faster is certainly progress, but it fits reality to treat it with the attitude that “being able to decide fast and cheap” is a different thing from “being able to decide correctly.”
The model has begun to be offered through Microsoft Foundry, and a rollout to OpenRouter has also been announced. OpenAI, Cloudflare, and others have likewise started building similar dedicated “score-the-options” models, and the trend of carving out decision-making alone is spreading across the industry.
References: Introducing Microsoft-Decision-1, our model for fast decision-making（Microsoft） ／ Microsoft launches Decision-1 model in Foundry（TestingCatalog） ／ Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model（MarkTechPost） ／ Microsoft puts a 9B decision model in Foundry for agent workflows（Runtimewire） ／ Microsoft-Decision-1 — API Pricing & Providers（OpenRouter）
