Microsoft has launched Microsoft-Decision-1, a small AI model designed to make decisions between predefined options instead of generating long-form responses.
The model is based on Alibaba’s open-weight Qwen3.5-9B and has been post-trained by Microsoft specifically for tasks such as classification, routing, verification, and workflow control.
Microsoft says Decision-1 can outperform or match much larger general-purpose AI models on decision-focused workloads while running substantially faster and at a fraction of the cost.
What Microsoft-Decision-1 Does
Decision-1 receives a fixed set of possible answers and assigns a calibrated probability score to each option.
It can handle yes/no questions, multiple-choice tasks, and ratings. Microsoft also designed it to evaluate AI responses or agent actions against a predefined rubric.
Potential applications include:
- Classification
- Routing
- Prioritization
- Verification
- Workflow control
- AI response grading
- Agent action evaluation
Unlike a conventional chatbot, Decision-1 is not designed to write essays, emails or conversational answers.
Microsoft-Decision-1 Specifications
| Specification | Microsoft-Decision-1 |
|---|---|
| Base Model | Qwen3.5-9B |
| Developer | Microsoft |
| Input | Text |
| Context Window | 32,000 tokens |
| Primary Function | Decision-making and classification |
| Output | Probability scores across predefined choices |
| Input Price | $0.042 per 1 million tokens |
| Output Price | Free |
| Availability | Microsoft Foundry, OpenRouter |
Microsoft warns that the model should not be the sole basis for high-impact decisions involving people, including areas such as employment or credit.
Decision-1 Tops Microsoft’s 36-Benchmark Test
Microsoft evaluated Decision-1 across 36 benchmarks containing nearly 150,000 questions that were kept blind from training.
According to Microsoft, Decision-1 achieved the highest overall accuracy among the models included in its evaluation.
The company also emphasized latency. Decision-1 was reportedly 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol.
Microsoft Decision Benchmark Results
| Model | Relative Speed vs Decision-1 | Key Result |
|---|---|---|
| Microsoft-Decision-1 | 1x baseline | Highest overall accuracy in Microsoft’s 36-benchmark evaluation |
| H2O-Lightning-4B v1.1 | Decision-1 was 2.5x faster | Runner-up in Microsoft’s comparison |
| GPT-6 Sol | Decision-1 was 35x faster | Much larger general-purpose model |
| Jev | Decision-1 was 2.8x faster | Competing decision model |
The results come from Microsoft’s own testing, so they should be interpreted within the company’s benchmark methodology rather than as independent measurements.
Only 1.3% Decision Change After Prompt Rewording
Microsoft also tested how consistently Decision-1 responds when the same question is phrased or ordered differently.
The model changed its decision only 1.3% of the time on average when Microsoft reworded or reordered equivalent requests.
That consistency is particularly important for tasks such as routing and automated workflows, where minor changes in phrasing should not produce completely different decisions.
| Test | Microsoft-Decision-1 Result |
|---|---|
| Benchmarks tested | 36 |
| Questions evaluated | Nearly 150,000 |
| Average decision change after rewording/reordering | 1.3% |
| Safety benchmarks | 11 |
| Safety requests | 5,250 |
Xbox Says It Was 14x Faster and 200x Cheaper Than GPT-6 Sol
Microsoft has already tested Decision-1 in internal products and teams.
Xbox Research used the model to categorize more than 10,000 pieces of player feedback and reviews into different themes.
Microsoft says Decision-1 delivered quality competitive with GPT-6 Sol while being more than 14 times faster and around 200 times cheaper for that workload.
Xbox Research Test
| Metric | Microsoft-Decision-1 vs GPT-6 Sol |
|---|---|
| Quality | Competitive |
| Speed | 14x+ faster |
| Cost | Up to 200x cheaper |
| Feedback Items Processed | 10,000+ |
This use case illustrates where specialized decision models could have an advantage. Sorting thousands of customer reviews does not necessarily require a large model capable of generating complex prose.
Copilot Test Shows 100x Speed Advantage
Microsoft’s Copilot team also evaluated Decision-1.
According to Microsoft, its decision quality was competitive with GPT-5.6 Luna, while Decision-1 was around 100 times faster.
| Internal Test | Quality Comparison | Speed Advantage |
|---|---|---|
| Xbox Research | Competitive with GPT-6 Sol | 14x+ faster |
| Microsoft Copilot | Competitive with GPT-5.6 Luna | 100x faster |
Safety Testing
Microsoft evaluated Decision-1 using 5,250 prompts across 11 safety benchmarks.
The evaluations covered areas including:
- Harmful content
- Jailbreak attempts
- Prompt injection
Decision models can be particularly important in AI agent systems because they may be used to approve, reject or route actions taken by another model.
Extremely Low Pricing
Microsoft has priced Decision-1 at just $0.042 per million input tokens, while output tokens are free.
| Token Type | Price per 1 Million Tokens |
|---|---|
| Input | $0.042 |
| Output | Free |
The unusual pricing reflects the model’s specialized purpose. Instead of producing hundreds or thousands of output tokens, it primarily returns scores for a predefined set of choices.
Microsoft Plans OpenAI and MAI-Based Versions
Decision-1 currently uses Qwen3.5-9B as its foundation, but Microsoft says it plans to rebase the system on other models.
Future versions could use OpenAI models as well as Microsoft’s own MAI models.
Microsoft has not explained why it intends to maintain Decision-1 across multiple model families.
Decision Models Are Becoming a Major AI Category
Microsoft is entering an increasingly crowded market. TypeSafe AI launched Jev several weeks before Decision-1, while OpenAI has placed its Decisions API into public beta.
Cloudflare, Strands, Liquid AI, Perplexity and Snowflake are also developing decision-focused AI systems. Microsoft says Decision-1 was 2.8 times faster than Jev in its latency benchmark.
The broader idea is straightforward: not every AI task requires a massive general-purpose model. For workloads where an AI only needs to choose the best option, route a request, or verify an action, smaller decision models could potentially deliver similar results with dramatically lower latency and cost.
The post Microsoft’s New AI Model Is 35x Faster and Up to 200x Cheaper Than GPT-6 Sol appeared first on ProPakistani.