Microsoft‑Decision‑1: Fast, Low‑Cost Decision‑Scoring Model – What It Is, How It Performs, and Why It Matters

TL;DR – Key Takeaways

What it is: A 9B‑parameter decoder model that outputs calibrated probabilities for each answer choice in a single forward pass.
Performance: 35× faster (≈35 ms vs 1.2 s) and up to 14× cheaper than GPT‑6 Sol, with accuracy comparable to Jev‑Bench‑1.0‑Large (≈98 % on the 36‑benchmark suite) and better than many larger LLMs.
Pricing: $0.042 USD per million input tokens; output tokens are free.
Availability: Microsoft Foundry (AI Azure catalog) and OpenRouter (model ID microsoft/microsoft-decision-1).
Use cases: Real‑time routing, safety filtering, agentic decision gating, data labeling, incident‑response triage, scientific‑discovery screening, and any workflow that needs fast, calibrated choice probabilities.
Future roadmap: Re‑training on MAI and OpenAI models, expanded benchmark coverage, and tighter confidence‑calibration.

1. What Is Microsoft‑Decision‑1?

Microsoft‑Decision‑1 is a decision‑scoring model designed to turn a set of answer options into a probability distribution in a single forward pass. Unlike generic LLMs that generate free‑form text and then require post‑processing to extract a score, Decision‑1 directly outputs a calibrated probability for each option, making it ideal for:

Routing (choose the best downstream model or service),
Safety/Compliance gating (accept/reject based on calibrated risk),
Agentic control (decide whether an autonomous agent should act),
Classification (labeling, sentiment, intent), and
Scientific screening (filter hypotheses, compounds, or experiments).

The model is fine‑tuned from Qwen‑3.5‑9B, a 9‑billion‑parameter open‑source decoder. Microsoft plans to re‑base it on MAI and OpenAI models later, but the current version already delivers state‑of‑the‑art speed‑cost trade‑offs.

API Contract (OpenRouter)

POST https://openrouter.ai/api/v1/chat/completions
{
  "model": "microsoft/microsoft-decision-1",
  "messages": [{"role": "user", "content": "Question: What is the capital of France? Options: A) Berlin B) Paris C) Madrid"}],
  "temperature": 0,
  "max_tokens": 1
}

The response contains a single token that maps to the selected option, and the logprob field gives the calibrated probability for each choice.

2. Benchmark Results – Speed, Cost, and Accuracy

2.1 Benchmark Suite

Microsoft evaluated Decision‑1 on 36 public benchmarks (including Jev‑Bench, Strands‑Decider, and a custom “cost‑vs‑accuracy” suite) covering:

Classification (e.g., intent, safety),
Routing (model selection),
Ranking, and
Decision‑making under constraints. All benchmarks were blinded to the model (training data excluded) and run on the same hardware for fair comparison.

2.2 Speed & Latency

Model	Median latency (p50)	Relative speed
Microsoft‑Decision‑1	≈35 ms	1× (baseline)
GPT‑6 Sol	~1.2 s	≈35× slower
Jev‑Bench‑1.0‑Large (Qwen‑1.0‑Large)	~400 ms	≈11× slower
deck‑31B (H2O‑Lightning‑4B)	~800 ms	≈23× slower

Decision‑1’s single‑pass architecture eliminates the need for multiple token generations, delivering the fastest scores among all evaluated models.

2.3 Cost

Input‑only pricing: $0.042 USD per million tokens (≈$0.000042 per 1 k tokens).
Outputs are free – a single token per decision, so the effective cost per decision is essentially the input token cost.
Compared to GPT‑6 Sol (≈$0.60 USD per million tokens) Decision‑1 is ~14× cheaper.

2.4 Accuracy & Calibration

Model	Overall accuracy (average across 36 benchmarks)
Microsoft‑Decision‑1	≈98 % (matched or exceeded Jev‑Bench‑1.0‑Large)
Jev‑Bench‑1.0‑Large	97 %
GPT‑6 Sol	94 %
deck‑31B	92 %

Calibration – Decision‑1’s probability scores are directly comparable across tasks, unlike generic LLMs where raw token probabilities can be poorly calibrated. In the “robustness” test (1.3 % flip rate across 8 perturbations), Decision‑1 flipped only 1.3 % of its decisions, the lowest among all models tested.

3. Real‑World Use Cases Demonstrated by Microsoft

Category	Example	How Decision‑1 Is Used
Agentic gating	An autonomous agent proposes to delete a user’s data. Decision‑1 evaluates the safety policy and returns a 99 % probability that the action is safe, allowing the system to proceed automatically.	Provides a fast, calibrated safety check before the agent acts.
Routing	A chatbot needs to pick the best LLM for a user query (cost‑ vs. quality‑sensitive). Decision‑1 scores candidate models and selects the cheapest model that meets a quality threshold.	Enables dynamic model selection in milliseconds.
Safety filtering	Content moderation pipeline classifies user‑generated text as safe/unsafe. Decision‑1 returns a probability; if > 0.9, the content is auto‑approved, otherwise flagged for review.	Reduces latency in high‑throughput moderation.
Data labeling	Large corpora of social‑media posts are labeled with sentiment or intent. Decision‑1 assigns probabilities for each label, allowing downstream systems to weight decisions by confidence.	Improves labeling quality while keeping costs low.
Incident‑response triage	An on‑call engineer receives an alert with a description. Decision‑1 classifies the incident type (e.g., network, database, security) and routes it to the appropriate team.	Cuts mean‑time‑to‑triage from minutes to seconds.
Scientific discovery	A chemistry pipeline generates candidate compounds. Decision‑1 evaluates each against a rubric (toxicity, synthesizability) and ranks them, allowing the lab to focus on the top‑scoring 5 %.	Accelerates hypothesis screening by ~4× compared to manual review.

4. Pricing Details

Input‑only cost: $0.042 USD per million tokens (≈$0.000042 per 1 k tokens).
Output tokens: Free – the model is designed to emit a single token per decision, so the entire cost of a decision is the cost of the input prompt.
Comparison: GPT‑6 Sol costs roughly $0.60 USD per million tokens, making Decision‑1 ~14× cheaper for the same workload.
Billing: Pricing is listed on the OpenRouter model page and on the Microsoft Foundry catalog entry.

5. Availability & How to Try It

Platform	Access URL	Authentication
Microsoft Foundry	https://ai.azure.com/catalog/models/Microsoft-Decision-1	Azure AD (Azure subscription)
OpenRouter	https://openrouter.ai/microsoft/microsoft-decision-1	API key (free tier available)

Both platforms expose a Chat Completion endpoint compatible with the OpenAI API schema. The model is already live; no waiting list is required.

6. Roadmap & Future Work

Re‑training on MAI & OpenAI backbones – to improve reasoning and multilingual support while preserving the low‑latency architecture.
Expanded benchmark suite – adding more safety‑critical and domain‑specific tasks (e.g., medical triage, financial compliance).
Confidence‑calibration improvements – fine‑tuning on larger calibration datasets to tighten probability reliability.
Tooling – SDKs for Python, JavaScript, and .NET that expose a simple score(options, prompt) helper.
Edge deployment – a lightweight ONNX version for on‑device inference (e.g., Windows Copilot, Azure IoT Edge).

7. Community & Criticisms (What the HN Thread Says)

Pricing comparison: Some commenters noted that Microsoft only compared cost to GPT‑6 Sol, not to Jev‑Bench‑1.0‑Large (which is ~4× cheaper than Luna). Microsoft’s $0.042 USD/M‑token is still significantly lower than any GPT‑6‑type offering, but the community expects a direct Jev cost comparison.
Model choice: The decision to start from Qwen‑3.5‑9B (a Chinese‑origin model) raised eyebrows. Microsoft’s response: “We will re‑base on MAI and OpenAI models soon,” indicating a plan to diversify the backbone.
API documentation: Early users reported missing OpenAPI specs. Both Foundry and OpenRouter now provide Swagger/OpenAPI JSON for easy integration.
Naming & branding: A few jokes about “Microsoft Decision‑1” sounding like a Windows‑style UI component (e.g., “Clippy”). The name reflects the model’s purpose (decision scoring) and follows Microsoft’s product‑line naming convention (e.g., Copilot, Designer).
Use‑case relevance: Some commenters wondered if the model is over‑engineered for simple classification tasks. Microsoft’s positioning emphasizes agentic AI—where fast, calibrated decisions gate downstream actions—making the low‑latency, low‑cost profile essential.

8. Quick Start Code Snippet (Python)

import openai

client = openai.OpenAI(base_url="https://openrouter.ai/api/v1", api_key="YOUR_KEY")

prompt = (
    "Question: Which team should handle a security incident? "
    "Options: A) Tech Support B) Security Team C) Customer Success"
)

response = client.chat.completions.create(
    model="microsoft/microsoft-decision-1",
    messages=[{"role": "user", "content": prompt}],
    temperature=0,
    max_tokens=1,
    logprobs=True,
)

# Extract probabilities for each option
logprobs = response.choices[0].logprobs.top_logprobs[0]
probs = {k: round(10**v * 100, 2) for k, v in logprobs.items()}
print(probs)  # e.g., {'A': 2.3, 'B': 94.7, 'C': 2.9}

The dictionary probs gives a calibrated probability for each answer choice, ready for downstream decision logic.

9. Bottom Line

Microsoft‑Decision‑1 fills a missing niche in the LLM ecosystem: a fast, cheap, calibrated decision scorer that works out‑of‑the‑box for routing, safety gating, and agentic control. Its performance on the 36‑benchmark suite shows state‑of‑the‑art accuracy while delivering 35× lower latency and up to 14× lower cost than comparable GPT‑6 offerings. The model is already usable via Azure Foundry and OpenRouter, and the roadmap promises broader backbones, tighter calibration, and edge deployment.

If you are building autonomous agents, real‑time routing pipelines, or high‑throughput safety filters, Microsoft‑Decision‑1 is now a practical, production‑ready option that can dramatically reduce both latency and spend.