Microsoft's Decision-1 model joins the rapidly expanding AI decision-model competition

Meta Description: Microsoft Decision-1, built on Qwen3.5-9B, leads 36 benchmarks with 83.5% accuracy and 85 ms latency, while input tokens cost $0.042 per million.
Tags: Microsoft Decision-1, Qwen3.5-9B, Microsoft Foundry, OpenRouter, Jev 1.13.0
Featured Image: abstract network of decision nodes and branching paths on a dark blue gradient, no text, no logos, no people

Microsoft now has a decision model of its own, called Decision-1, and it enters a category already occupied by OpenAI, Cloudflare, and the "original" from startup Jev, which started the decision model trend in mid-September. The pitch is fast, structured choices: classifications, evaluations, and routing decisions. Microsoft says Decision-1 has "the potential" to "guide and control agents through complex environments." The model matters because smaller, task-specific systems can be cheaper and more predictable than general-purpose language models when an application needs to choose an action, sort inputs, or send requests to the right tool.

Background: A Decision-Model Category Forms

Decision models are narrow systems built to make repeated, bounded choices instead of producing long open-ended answers. The timeline starts in mid-September, when Jev introduced the trend. The concept spread quickly, and larger AI infrastructure players moved into the same space. OpenAI and Cloudflare became associated with similar approaches, while Cloudflare also offers open-source Clef models based on Qwen. Microsoft's launch therefore follows a pattern that had already developed over several weeks, rather than creating the category from zero.

Development: Qwen3.5-9B Becomes a Benchmark Claim

Decision-1 is based on Qwen3.5-9B. In Microsoft's comparison table, Decision-1 sits at the top: 83.5% accuracy and 85 ms latency, ahead of Jev 1.13.0. Microsoft calls it the most accurate model it tested across 36 benchmarks, a set that covers nearly 150,000 questions. For speed, Microsoft reports a 2.5 times advantage over H2O-Lightning-4B, the runner-up. The chart is credited to Microsoft. Cloudflare's open-source Clef models were also built on Qwen, but they were not part of this comparison, a detail that keeps the published ranking narrower than the broader Qwen-based model landscape.

Latest Progress: Distribution and Pricing Put Pressure on Jev

Decision-1 is available through Microsoft Foundry and OpenRouter. Input tokens cost $0.042 per million, and output tokens are free. That pricing is useful for high-volume filtering or routing, where a system reads many tokens but returns only a short label, score, or route. It also creates pressure on Jev, which likely had a rough few weeks after the concept spread quickly. The technology was quickly adapted and surpassed using open small language models, and Microsoft's entry adds a major vendor with distribution channels and a benchmark table.

What to Watch Next: Benchmark Scope and Open-Model Speed

The next stage of the decision model category will be tested by how well specialized systems hold up when open small language models are reused, tuned, and distributed quickly. Cloudflare's Clef models show that Qwen-based decision systems are not limited to one vendor, even if they were excluded from Microsoft's comparison. For readers, the current signal is that the race has moved from concept to deployment: the relevant questions are no longer whether decision models can make structured choices, but which accuracy, latency, price, and platform access combination will become the default for agent routing and classification.