Table of Contents
Microsoft has released Decision-1, a compact AI model whose job is to pick between options rather than write long answers. The launch was revealed on Friday, October 9, and the target is applications that have to settle numerous minor questions on the way to completing an automated job.
Microsoft’s own tests suggest Decision-1 beats rival models on accuracy while responding far more quickly. It is available through Microsoft Foundry, with a listing on OpenRouter planned. Teams building AI agents could use it to hold down their operating expenses.
What Is Microsoft Decision-1 AI?
Most language models generate free-form text for almost any prompt. Decision-1 does something narrower: it weighs a set of alternatives and returns the best one. It attaches calibrated probability scores to each option, so an application can act on the output straight away.
Supported formats include binary yes/no queries, picking from a list of answers, and assigning scores or ratings. It can also grade AI-written answers or agent actions against criteria a developer defines.
Picture an agent that keeps asking what to do next. It might use Decision-1 to choose which model to call, retry a failed step, or hand a tricky case to a human worker. Microsoft pitches the model as an inexpensive decision layer alongside bigger systems, so developers need not pay for a heavyweight general model every time a minor judgement is required.
How Decision-1 Was Built
Microsoft started from Qwen3.5-9B, an open-weight model from Alibaba, and trained it further for this purpose. The result scores candidate answers in a single pass, which keeps response times short.
Microsoft also intends to release variants built on different foundations, such as its in-house MAI models and models from OpenAI.
CEO Satya Nadella announced the launch and pointed to the model’s results on structured decision tasks. He added that teams inside Microsoft are trying it for incident response, quality control and scientific discovery.
Achint Srivastava, a vice president in the Office of the CTO, explained why delays matter. If each of 20 sequential decisions runs 100 milliseconds slower, the workflow loses a full two seconds, a real drag when many steps depend on earlier ones.
Benchmark Results and Performance
Microsoft says Decision-1 posted the top accuracy across 36 benchmarks, covering nearly 150,000 questions held out of training, and was the quickest model in the comparison.
On median latency, the company reports it ran 4.5 times faster than Quyet-1.0-Large and roughly 35 times faster than OpenAI’s GPT-6 Sol.
To check consistency, Microsoft reworded questions and reordered options in eight different ways. Decision-1 changed its answer only 1.3% of the time on average. Separately, the company ran safety checks using 5,250 prompts spread over 11 benchmarks.
Every figure here comes from Microsoft’s own testing and has not been independently confirmed. Results can shift with the task, evaluation method and workload, so third-party testing will be needed to show whether the advantages hold elsewhere.
Internal Trials and Business Uses
Several Microsoft teams tried the model on real work:
- Xbox Research grouped upward of 10,000 player comments into themes. Quality was reportedly similar to GPT-6 Sol, with the job running more than 14 times faster at about one two-hundredth of the cost.
- The Copilot team used it to judge AI-generated replies. Microsoft says it rivalled GPT-5.6 Luna while grading 100 times faster.
- Microsoft Discovery trialled Decision-1 for scoring in science-related workflows, where the company says its scores were 46 times more consistent than those from an LLM-based method.
These trials point to uses in feedback analysis, response grading and research, though they are internal findings rather than verified comparisons.
Pricing and Availability
The pricing is set low so that making decisions often does not become expensive.
| Feature | Details |
| Model | Microsoft Decision-1 |
| Base model | Qwen3.5-9B |
| Input cost | $0.042 per million tokens |
| Output cost | Free |
| Available now | Microsoft Foundry |
| Planned | OpenRouter |
| Main purpose | Structured decision-making |
Low prices could attract organisations whose agents check options constantly, letting developers add extra safeguards without sharply raising the bill.
The launch also fits the rising interest in agentic AI, where tasks are completed through chains of connected steps. In such setups, small specialised models can take on narrow jobs while larger ones handle complex reasoning and writing.
Market data cited by Futu showed Microsoft’s shares rising after the announcement, though that move alone says little about how investors view the model’s long-term commercial value.
FAQs
1. What is Microsoft Decision-1 AI?
A specialised model that chooses among predefined options and scores each with a probability, supporting fast decisions inside software.
2. How much does it cost?
Input is priced at $0.042 per million tokens. Output tokens are free.
3. Where can I use it?
Through Microsoft Foundry now, with OpenRouter access planned.
4. What can it do?
Classification, multiple-choice selection, ratings, grading of AI responses and decision steps within agent workflows.
5. Are the performance claims independently verified?
No. The benchmarks and internal trial results are all reported by Microsoft itself.
Conclusion
Microsoft Decision-1 takes a focused approach to cutting the time and cost of repeated AI decisions. Its structured outputs, probability scoring and low input pricing could make it useful for autonomous agents and automated business tools.
The reported numbers are promising, but independent testing will show how well they hold up. As AI agents spread, specialised decision models like this one may become a standard piece of faster, more consistent and cheaper automated workflows.