Microsoft has released Microsoft-Decision-1, a specialized decision-scoring model post-trained from Alibaba’s Qwen3.5-9B. Available on Microsoft Foundry and OpenRouter, the model evaluates input against a closed set of options and returns calibrated probabilities in a single forward pass, entirely bypassing text generation to support direct software branching.

In internal testing across 36 benchmarks, Microsoft reported an average accuracy of 83.5% with median latencies of 85 ms, outperforming systems like Quyet-1.0-Large and GPT-6 Sol. However, the evaluation methodology has drawn pushback from competitors like H2O.ai, which disputed Microsoft’s adjusted latency figures and noted its own models were omitted from the comparison charts.

Microsoft-Decision-1 is priced at $0.042 per million input tokens with no charge for output tokens on OpenRouter’s Decisions API. Microsoft indicated that future iterations of its decision-scoring architecture will be rebased on MAI and OpenAI base models.

Why it matters

  • Decision models trade text generation for sub-100ms structured probability outputs, significantly lowering latency for automated agent routing and control.

  • Zero token generation fees and competitive input rates ($0.042/1M) drastically lower operational overhead for continuous workflow branching.

  • Developers using standard chat completion SDKs must adopt specialized APIs (like OpenRouter Decisions API or Direct from Azure) to integrate option-scoring models.

Source: marktechpost.com