Skip to main content

Microsoft's New AI Model Nails 36 Tests, Slashes Decision-Making Costs

Microsoft's New AI Model Nails 36 Tests, Slashes Decision-Making Costs

Microsoft has quietly rolled out a new AI model that's turning heads in the tech world. Dubbed Microsoft-Decision-1, it's built for structured decision-making tasks, and early results are impressive. In 36 rigorous benchmark tests covering roughly 150,000 questions, it consistently came out on top for accuracy—all while being blazing fast. Think of it as a speed-demon that doesn't sacrifice precision.

What It Does

At its core, this model is a pro at picking the best option from a predefined list. It doesn't just choose; it assigns probability scores to each alternative, making it perfect for real-world business scenarios like routing requests or classifying content. You can plug it into existing apps, AI agents, or workflow systems to streamline complex processes. No fuss, just efficient decision-making.

Under the Hood

Microsoft-Decision-1 is built on Qwen3.5-9B and underwent specialized post-training. Microsoft hints that it'll migrate to other base models down the line. But what about reliability? The model holds up well under pressure—it's robust against perturbations, with a low decision reversal rate, and it passed multiple security tests with flying colors. So you can trust it when the stakes are high.

The Price Tag

Here's where it gets really interesting: the cost. At just $0.042 per million input tokens, with output completely free, it's a bargain. In tests, it matched the quality of some larger models on similar tasks but blew them away on speed and cost. That's a game-changer for businesses watching their budgets.

Where to Get It

The model is now available to developers via Microsoft's AI development platform. And soon, it'll be integrated into OpenRouter, giving devs even easier access. If you're building AI-driven workflows, this is one tool you'll want to check out.

Key Points:

  • Microsoft-Decision-1 excels in structured decision-making, topping 36 benchmarks.
  • It selects optimal solutions and assigns probability scores, ideal for routing and classification.
  • Built on Qwen3.5-9B, with plans to migrate to other base models.
  • Extremely cost-effective: $0.042 per million input tokens, free output.
  • Available now on Microsoft's AI platform, coming soon to OpenRouter.