Microsoft has entered the decision-model segment with Decision-1, a model built for fast and structured choices inside AI systems. The company reports top accuracy across 36 benchmarks covering nearly 150,000 questions and speed 2.5 times higher than runner-up H2O-Lightning-4B. For business, this matters because routing and classification are becoming a separate paid layer in agent architectures.
How Microsoft positions Decision-1
Decision-1 is designed for classifications, evaluations, and routing decisions, rather than open-ended generation or conversation. Microsoft says it has the potential to guide and control agents through complex environments, placing it as a control component inside larger workflows. The model is based on Qwen3.5-9B, continuing the pattern of building decision models on open small language models. Availability is already defined: access runs through Microsoft Foundry and OpenRouter.
Pricing follows an unusual structure for this class of models. Input tokens cost $0.042 per million, while output tokens are free, which fits short structured answers rather than long text. The accuracy claim rests on a broad test set of 36 benchmarks and nearly 150,000 questions, according to Microsoft. On speed, the stated gap is large: 2.5 times faster than H2O-Lightning-4B. Cloudflare open-source Clef models, also based on Qwen, were not included in that comparison.
The launch follows a rapid sequence of similar releases. Jev started the decision-model trend in mid-September, and models from OpenAI and Cloudflare have since appeared alongside Microsoft entry. The source notes that Jev has likely had difficult weeks as the idea spread and was adapted by others using open small language models. In other words, a startup concept turned within weeks into a competitive category with large-platform distribution. That speed explains why comparisons now focus on benchmarks, latency, and price.
What this means for agent architectures
For companies running agents, a dedicated decision model can separate routine triage from expensive reasoning. Classification of requests, quality evaluation, and routing to the right tool or agent become a low-cost operation at $0.042 per million input tokens with no output charge. Small firms gain a ready component through OpenRouter without managing infrastructure, while large firms can standardize routing inside Microsoft Foundry. The practical effect is lower cost per decision and more predictable behavior in multi-step processes.
The limits deserve the same attention as the headline figures. The accuracy leadership and speed advantage come from Microsoft testing, and Clef was excluded, so buyers cannot treat the ranking as independent. Performance across 36 benchmarks does not guarantee the same result on company-specific categories, thresholds, and edge cases. It would be prudent to test Decision-1 against H2O-Lightning-4B, Clef, and current rules on own data, checking latency, error types, and behavior under changed instructions.
The marker to watch is whether Foundry and OpenRouter usage turns Decision-1 into a default router for production agents. Wider adoption, published customer cases, and independent benchmark results would confirm the category. If rival vendors respond with lower prices or broader comparisons, decision logic will become commodity infrastructure. Then competition will move from model scores to governance, monitoring, and control of agents.
