Microsoft-Decision-1: Ranking, Pricing and Test Results

👁 113 reads

Microsoft-Decision-1 is a decision-making model from Microsoft, released on October 9, 2026, in Microsoft Foundry. It does not write text: it takes text or JSON and questions with a fixed set of answer choices, then returns probabilities. In the Artificial Analysis Decision Index ranking dated October 13, 2026, it scores 53.8 points (third out of 11). Among decision models, it is the most likely to admit when it does not know an answer, and it is among the most cost-effective: one million tasks cost $30.7. In my Polish-language test, it routed 44 of 47 customer requests to the correct department. I cover the category in my guide to AI decision models.

What Microsoft announced

Achint Srivastava, Microsoft’s corporate vice president of software engineering in the Office of the CTO, introduced the model on October 9. Microsoft describes it as a model for routing, classification, prioritization, verification and process control. The company makes the following claims:

• Highest accuracy in Microsoft’s comparison across 36 test sets (nearly 150,000 questions).
• 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol.
• It changes its decision in an average of 1.3% of cases after minor changes to the text, and never after option descriptions are reworded.
• It refused harmful behavior in 5,250 attempts across 11 safety test sets.

These figures come from Microsoft’s own evaluation. Artificial Analysis conducted an independent assessment.

Artificial Analysis benchmark results

• Best at saying “I don’t know.” When it does not know the answer, it chooses “I don’t know” rather than guessing 40% of the time, the best result among the decision models. Its overall knowledge accuracy is the lowest in this group, however, at 28.2.
• Strong at tool selection - 81.9, the third-best score after GLiDE and Celeris-1, and well ahead of Claude Haiku 5.5 low (72.6).
• Weaker on safety classification - 88.2 is one of the lower scores on WildGuardTest.
• Among the most cost-effective. Artificial Analysis groups it with Jev and Decider as a model for which a higher score comes at a higher price.

My Polish-language test (October 13, 2026)

Through OpenRouter, I gave it the same test set as eight other decision models and two LLMs: 50 Polish customer requests for a store (routing to one of seven departments and assessing urgency), plus 10 conversations in which an assistant had to choose whether to call a tool, ask a follow-up question, answer directly or refuse. The test was conducted in Polish. Any translated examples of the test material should be understood as translations of the original Polish prompts, not as prompts retested in another language.

• Department routing: 44 out of 47 - on par with the leading decision models and one request behind Claude Haiku 5.5.
• Urgency: 42 out of 50 - its weakest result in the test. At a 0.5 threshold, it marked eight cases as urgent when they could have waited (for example, a package stolen from a doorstep, with 1.0 confidence). You would need to raise the threshold using your own data.
• Confidence: At confidence of at least 0.9, it routed 40 of 43 requests to the correct department.
• Tokens: Polish text used an average of 273 tokens per task, the fewest among the decision models in the test (Jev: 637). One million such tasks would cost about $11.5.
• Latency: Median of 443 ms from Poland through OpenRouter (Claude Haiku 5.5: 1,689 ms).

Access and data processing in the EU

• Microsoft Foundry: Requires an Azure subscription with a payment method and a Foundry project. Endpoint: <foundry-resource>/providers/microsoft/v1/systemone, with a format compatible with Jev.
• Deployment types: GlobalStandard (processing in any Azure region) or DataZoneStandard in selected regions. With the latter, processing stays within the data zone, such as the EU. Microsoft does not list all regions that offer this model on its model page.
• OpenRouter: $0.042, with Azure as the provider. It worked from Poland without an Azure account when I checked on October 13.

Microsoft warns in its documentation that results may depend on wording and option order, calibration works best on familiar task types, and for decisions about people (credit, employment or health), the model should only support a human decision-maker.

When to choose Microsoft-Decision-1

• Consider it for: gates and filters where the model should say “I don’t know” rather than guess; tool selection in an agent; companies already using Azure that want processing in the EU.
• Probably not for: content moderation (GLiDE, Decider and Jev perform better), questions requiring broad knowledge, or the lowest possible cost (Decider v1.1 27B is almost twice as cheap).

Frequently asked questions

How much does Microsoft-Decision-1 cost?

$0.042 per million input tokens; output is free. The price is the same in Microsoft Foundry and on OpenRouter (as of October 13, 2026).

What model is it based on?

Microsoft says it is Qwen3.5-9B with additional post-training, or 9 billion parameters.

Does it work in Polish?

Yes. In my Polish-language test, it routed 44 of 47 customer requests to the correct department and chose the right tool in 10 out of 10 conversations.

Sources

• Microsoft: Microsoft-Decision-1 in Foundry (October 9, 2026)
• Microsoft Learn: Deploy and use Microsoft-Decision-1 (checked October 13, 2026)
• OpenRouter: microsoft/microsoft-decision-1
• Artificial Analysis Decision Index v1.0 (October 13, 2026)
• My test: October 13, 2026, 60 tasks through OpenRouter from Poland.

// read moreWhich AI Model for Work in 2026? Choose by Task, Not RankingGuidesMercury Decide vs Celeris-1: Speed and BenchmarksBenchmarksMAI-Image-2.6 Review: Pricing, Availability & Test ResultsBenchmarks

Kinetyra