Microsoft releases Decision-1, a Qwen3.5-9B model for fast structured decisions
Microsoft has released Decision-1, its own decision model, joining a trend that already includes decision models from OpenAI, Cloudflare and the startup Jev, whose model The Decoder calls the "original". Decision-1 is built for fast, structured decisions. According to Microsoft, it handles classifications, evaluations and routing decisions, and it has "the potential" to "guide and control agents through complex environments."
The model is based on Qwen3.5-9B. Microsoft says it is the most accurate model tested across 36 benchmarks covering nearly 150,000 questions. On speed, Microsoft says it is 2.5 times faster than the runner-up, H2O-Lightning-4B. Cloudflare's open-source Clef models, which are also based on Qwen, were not included in the comparison.
Decision-1 is available through Microsoft Foundry and OpenRouter. Input tokens cost $0.042 per million, and output tokens are free.
The Decoder places the launch in a short history. Jev kicked off the decision model trend in mid-September and, in the article's words, has likely had a rough few weeks. The idea caught on fast, but the tech was quickly adapted and surpassed using open small language models.
Key facts
- Microsoft's Decision-1 is built on Qwen3.5-9B and aimed at fast, structured decisions: classification, evaluation and routing.
- Microsoft says it is the most accurate model tested across 36 benchmarks covering nearly 150,000 questions, and 2.5 times faster than the runner-up, H2O-Lightning-4B.
- Available through Microsoft Foundry and OpenRouter; input tokens cost $0.042 per million and output tokens are free.
- Cloudflare's Qwen-based, open-source Clef models were not included in the comparison.
- Jev started the decision model trend in mid-September; The Decoder notes the tech was quickly adapted and surpassed with open small language models.
Why it matters
Decision models are a new niche: compact models tuned to make quick, structured calls rather than write long answers. Jev kicked the idea off in mid-September, and OpenAI, Cloudflare and now Microsoft have their own versions. Microsoft's entry is notable because it is built on an existing Qwen3.5-9B base and priced at $0.042 per million input tokens with free output tokens. The Decoder's reading is that the idea caught on fast but was quickly adapted and surpassed with open small language models.
Who it affects
Developers who need a model for classification, evaluation or routing decisions, the tasks Microsoft says Decision-1 handles. Microsoft pitches it as having the potential to guide and control agents, so teams building agent systems are the obvious audience. It also puts pressure on other decision model makers named in the article: OpenAI, Cloudflare and Jev.
How to use it
Decision-1 is available through Microsoft Foundry and OpenRouter. Input tokens cost $0.042 per million, and output tokens are free. The source does not say whether Decision-1 is open-weights or under what licence, and no release date is given.
How solid is it
The figures are Microsoft's own: the most accurate model tested across 36 benchmarks covering nearly 150,000 questions, and 2.5 times faster than the runner-up, H2O-Lightning-4B. The accuracy claim is Microsoft's own, with no independent verification mentioned and no accuracy scores given. The article is secondary coverage of a first-party launch and does not name the 36 benchmarks.
Risks and caveats
The comparison leaves out Cloudflare's open-source Clef models, which are also based on Qwen, so the accuracy and speed claims do not cover that direct rival. The source does not say how the speed comparison was measured (hardware, setup) or give absolute speed figures. It does not say how the Decision-1 pricing compares with competitors. The Decoder's own point is that the decision model idea has already been surpassed by open small language models once, which is a reminder that the lead in this category may not last.
“the potential" to "guide and control agents through complex environments.”
— Microsoft, as reported by The Decoder