Artificial Intelligence
Microsoft launches Microsoft-Decision-1 model
Microsoft said it had released Microsoft-Decision-1, a decision-scoring model for structured tasks, and claimed it outperformed competing models on its benchmark tests and internal evaluations.
- Microsoft said it had introduced Microsoft-Decision-1 and made it available in Microsoft Foundry and through OpenRouter.
- The company said it had post-trained the model from Qwen3.5-9B and would later rebase it on other models, including Microsoft AI (MAI) and OpenAI.
- Microsoft said its comparison covered 36 benchmarks and nearly 150,000 questions, and that the model achieved the highest accuracy in that set.
- Microsoft said the model was 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol in its measurements.
- The post said XBOX Research, the Copilot team and Microsoft Discovery had used or tested the model internally, and pricing was set at $0.042 per million input tokens with output tokens free.
Microsoft said it had introduced Microsoft-Decision-1 as a model for fast decision scoring, and it said the system was available in Microsoft Foundry and OpenRouter. The company described it as a tool for routing, classification, prioritisation, verification and workflow control.
Microsoft said it had post-trained the model from Qwen3.5-9B and planned to rebase it later on other models, including Microsoft AI (MAI) and OpenAI. It said the model returned calibrated probability scores for fixed answer choices and could handle yes-or-no, multiple-choice and rating tasks.
In its benchmark comparison, Microsoft said it had tested the model across 36 benchmarks covering almost 150,000 questions. The company said it recorded the highest accuracy in that set, and that the model was 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol.
The post also said XBOX Research had used the model on more than 10,000 pieces of feedback and reviews, and that the Copilot team had tested it for chat and agent response quality. Microsoft said Microsoft Discovery had used it for adaptive replanning, and the post listed pricing at $0.042 per million input tokens, with output tokens free.
Named in this story
Companies
- Microsoft
- introduced the model and made its claims
- OpenAI
- named as another rebasing target for the model
Organisations
- Microsoft AI (MAI)
- named as a later rebasing target for the model
- XBOX Research
- used the model to process feedback and reviews
- Copilot team
- tested the model for chat and agent response quality
Products and systems
- Microsoft-Decision-1
- the new decision-scoring model being announced
- Microsoft Foundry
- one of the places where the model was made available
- OpenRouter
- another service offering access to the model
- Qwen3.5-9B
- the model Microsoft said it post-trained the system from
- H2O-Lightning-4B v1.1
- the benchmark runner-up in Microsoft’s comparison
- GPT-6 Sol
- the model Microsoft used as a comparison point
- Microsoft Discovery
- used the model in adaptive replanning tests
How the source tells it
The piece is openly promotional, with launch-day boosterism built from first-person claims, benchmark superlatives and repeated assertions of speed and broad usefulness.
- hype launch language and claims that the model opens up many use cases
- vendor boosterism the company repeatedly praised its own model and cited internal examples of success
- triumph superlative benchmark claims and speed comparisons cast it as the best and quickest