Microsoft-Decision-1
Microsoft released Microsoft-Decision-1, a decision-scoring model that returns calibrated probabilities for fixed answer options, in Foundry at $0.042 per million input tokens, with OpenRouter support coming soon.
Microsoft-Decision-1 does not write open-ended text. It takes a fixed set of options and returns a calibrated probability for each in a single structured call, for routing, classification, verification and agent control. Microsoft post-trained Qwen3.5-9B for this, and says it had the highest accuracy in a 36-benchmark comparison of nearly 150,000 questions held out from training. Its P50 latency is about 35 times faster than GPT-6 Sol.
- Date
- Friday 9 October 2026
- Lab
- Microsoft
- Kind
- model
- Access
- closed API
- Price
- $0.042 per million input tokens, output tokens free
Figures
| Measure | Value | Measured by |
|---|---|---|
| Benchmark accuracy rank | 1st across 36 benchmarks, nearly 150,000 questions benchmarks kept blind from training | company |
| P50 latency vs GPT-6 Sol | ~35x faster Microsoft's own benchmarking | company |
| Speed vs runner-up Quyet-1.0-Large | 4.5x faster Microsoft's tests, as reported by TestingCatalog | company |
| Xbox Research labeling vs GPT-6 Sol | over 14x faster, about 200x cheaper over 10,000 feedback items, internal trial | company |
| Decision flips across 8 perturbations | 1.3% on average reported by TestingCatalog from Microsoft's data | company |
All figures are Microsoft's own; the weights are not stated to be released, and the post says Microsoft plans to rebase the model on other models.
Sources
- commandline.microsoft.com/microsoft-decision-1-model-foundry/
- testingcatalog.com/microsoft-launches-decision-1-model-in-foundry/
- cryptobriefing.com/microsoft-launches-decision-1-ai-model/
- finance.biggo.com/news/4941336a-143f-4f17-9a27-3c8cdeb1c9a1
This record was checked and corrected against its sources on 10 October 2026. How we check