Microsoft launches Decision-1 model on Qwen3.5-9B base

Microsoft announced Microsoft-Decision-1 on October 9, 2026, introducing a specialized model designed exclusively for high-volume decision-making tasks like routing, classification, and verification. According to the launch details reported by Shattered, the model does not generate open-ended text; instead, it performs a single forward pass to score a fixed list of predefined answer options, returning a calibrated probability distribution for each. This narrow focus distinguishes it from general-purpose assistants like GPT-6 Sol or Claude Opus 5.5, positioning it as a tool for engineering teams building agent pipelines and automated verification systems.

The most significant technical detail in the launch is the model’s foundation. Microsoft-Decision-1 is post-trained from Qwen3.5-9B, a 9-billion-parameter open-weight model developed by Alibaba. The Register highlighted this choice, noting that Microsoft is leveraging a Chinese AI lab’s open-weight release to accelerate time-to-market. Microsoft has explicitly stated that this is a temporary architectural decision, with plans to rebase the model on its own Microsoft AI (MAI) and OpenAI models in the future. This admission underscores a pragmatic approach to product delivery, prioritizing immediate utility over long-term architectural purity for this specific use case.

Performance and Pricing Claims

Microsoft provided specific performance metrics in its launch materials, claiming Decision-1 achieved 83.5% accuracy across a comparison spanning 36 benchmarks and approximately 150,000 held-out questions. The company also reported significant speed advantages, stating the model runs 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol in terms of P50 latency. These figures are vendor-reported and have not yet been independently verified by third parties, a point of caution for developers evaluating the claims. Additionally, Microsoft cited a 98.7% perturbation-stability rate, indicating the model’s decisions remain consistent even with minor input rewordings, a critical factor for production systems relying on automated binary or multi-option calls.

Pricing for the model is available through OpenRouter, listed at $0.042 per million input tokens with free output. This pricing structure reflects the model’s nature, as its output consists of short probability scores rather than generated text, eliminating the need for output token metering. This cost profile positions Decision-1 as a highly economical option for high-frequency, low-complexity tasks, similar to the pricing strategies seen with Anthropic’s Claude Haiku.

Context in the AI Landscape

The launch of Decision-1 represents a shift in focus from general-purpose generation to specialized, high-efficiency decision-making. While the industry has largely concentrated on increasing model size and context length, Microsoft is targeting the specific needs of production systems that require fast, reliable, and cheap classification. This approach aligns with trends seen in other specialized models, such as JetBrains’ Mellum2.1, which prioritized compact parameter counts for specific coding tasks. Microsoft’s use of Qwen3.5-9B as a base is notable given the company’s own frontier model ambitions, suggesting a strategic willingness to utilize open-weight models from competitors when they offer a clear advantage in speed and cost for narrow applications.

What It Means for Developers

For developers, Microsoft-Decision-1 offers a streamlined solution for tasks that traditionally require parsing and interpreting free-text responses from larger models. By providing a direct probability distribution over predefined options, the model eliminates the need for complex parsing logic and reduces the risk of formatting errors. This is particularly useful for support ticket routing, content moderation, and agent output verification. However, developers should be aware that the model is currently in public preview on Microsoft Foundry, and the underlying base model is expected to change. Teams should plan for potential migration or re-evaluation when Microsoft rebases the model on its own architectures. The lack of independent benchmark verification also means that performance claims should be tested in specific production environments before full-scale deployment.

What to Watch

• Rebasing Timeline: Microsoft stated it will "soon" rebase Decision-1 on MAI and OpenAI models. The timing and impact of this transition on performance and pricing will be critical to monitor.
• Independent Benchmarks: Third-party verification of the 83.5% accuracy and 35x latency claims will provide a more reliable basis for adoption decisions.
• Stability in Production: Real-world testing of the 98.7% perturbation-stability claim under varying input conditions will determine the model’s reliability for automated systems.
• Pricing Changes: As the model moves out of preview and potentially rebases, pricing structures may adjust, affecting its cost-effectiveness compared to other specialized models.