The Decision 1 model provides ultra-fast speed in making calculated logical decisions to support AI agents and reduce response time in complex programming operations.
- The model’s philosophy and mechanism for settling specific choices
- Custom training and the development of a low-cost control layer
- Benchmark test results and response time superiority
- Successful field applications within Microsoft’s ecosystem
- Frequently asked questions
The model’s philosophy and mechanism for settling specific choices
Microsoft announced on Friday, October 9, the launch of a revolutionary artificial intelligence model named “Decision 1”. This new system differs radically from traditional language models, as it does not write open-ended creative texts or formulate lengthy conversational paragraphs. Instead, its core role focuses on precise selection among pre-determined options, assigning a calibrated probability value to each option, which allows computer software and automated systems to take direct actions and execute immediately without any delay.
The model has become available for immediate use via the Microsoft Foundry platform, and is scheduled to be available soon through the OpenRouter gateway. The company emphasized that “Decision 1” clearly outperformed both large language models and specialized decision-making models in both speed and accuracy standards. The company introduced this model at a highly competitive price of $0.042 per million input tokens, while making output tokens completely free, in a move aimed at encouraging developers to integrate it into complex systems.
Custom training and the development of a low-cost control layer
Microsoft introduced the model to act as a low-cost and reliable control and routing layer for automated AI agents. These agents require making dozens of instantaneous decisions between each step, such as determining the appropriate model to call, making a decision to retry upon failure, or determining the appropriate moment to transfer a transaction to a human employee. Each of these steps usually leads to an accumulation of time delays and high financial costs.
Satya Nadella, Chairman and CEO of Microsoft, revealed in a post on the X platform that the model achieves the highest levels of performance in structured decision-making tasks, noting that the company has already begun testing and applying it internally in multiple fields ranging from rapid response to technical faults and accidents and quality control to complex scientific discoveries. The company developed this model through additional, intensive training of Alibaba Group’s open-weights Qwen 3.5-9B model to evaluate alternatives and calculate probabilities in a single computing pass, with plans to rebuild it in the future on top of its own foundational models and OpenAI models.
Benchmark test results and response time superiority
Achyant Srivastava, Vice President in the Office of the CTO at Microsoft, explained the profound impact of speed in automated workflows, stating that adding 100 milliseconds to each decision in a sequence consisting of 20 consecutive steps leads to wasting two full seconds in processing a single transaction.
The results of comprehensive evaluations conducted by the company across 36 benchmarks comprising nearly 150,000 problems withheld from the training phases showed that “Decision 1” achieved the highest accuracy rate compared to all competitors. It also ranked first among models in terms of response speed, outperforming its closest runner-up, the Qwen 1.0 Large model, by 4.5 times, and OpenAI’s GPT-6 Soul model by about 35 times in average response time. Stability tests, which included rephrasing and reordering questions in eight different ways, showed that the model’s answer changed by only 1.3 percent on average.
Successful field applications within Microsoft’s ecosystem
Real-world experiments within the company’s sectors proved the viability of relying on this innovation. The Xbox gaming platform research team employed the model to sort and analyze more than 10,000 player feedback entries and classify them within specific axes, achieving quality parallel to the GPT-6 Soul model at 14 times the speed and at 200 times lower cost.
In the same context, the Copilot assistant team tested the model’s performance and found that it delivers competitive performance matching the GPT-5.6 Luna model while being 100 times faster in evaluating the quality of responses and outputs. Inside the Microsoft Discovery platform, the model outperformed in result consistency by 46 times compared to general language models, prompting the company’s shares to rise in financial market trading following this announcement.
Frequently asked questions
Question: What distinguishes the Decision 1 model from traditional language models?
Answer: The model does not generate creative texts; rather, it precisely chooses from specific options and assigns numerical probabilities to them to execute programming decisions immediately and in the shortest possible time.
Question: What is the prescribed financial cost for using this model?
Answer: The usage price is $0.042 per million input tokens, while output tokens are available completely free of charge.
Question: What was the base model that Microsoft relied on in developing this system?
Answer: The company relied on Alibaba’s open-weights Qwen 3.5-9B model, conducting specialized training to execute tasks in a single pass.
Question: What is the practical benefit of the model in supporting AI agents?
Answer: It acts as a fast and inexpensive control layer to settle repetitive decisions such as command routing, retries, and accurately evaluating output quality.
