Microsoft releases Decision-1, a Qwen3.5-9B-based model for classification and routing, at $0.042 per 1M input tokens

Microsoft has released Decision-1, a small model built on Qwen3.5-9B for fast, structured decisions. It is priced at $0.042 per 1M input tokens, with output tokens free, and is available through Microsoft Foundry and OpenRouter.

What Decision-1 does

Decision-1 targets three task types: classification, evaluation, and routing decisions. According to Microsoft, the model has "the potential" to "guide and control agents through complex environments." That agent-control use is framed as potential, not a demonstrated capability.

The model joins a category that has grown quickly since mid-September. Decision models from OpenAI and Cloudflare are already on the market, along with the "original" from startup Jev. All are built for fast, structured outputs rather than open-ended generation.

Specifications

Item	Detail
Base model	Qwen3.5-9B (about 9B parameters)
Input price	$0.042 per 1M tokens
Output price	Free ($0)
Availability	Microsoft Foundry, OpenRouter
Context window	Not yet disclosed
Training cutoff	Not yet disclosed

Benchmark claims

Microsoft says Decision-1 is the most accurate model tested across 36 benchmarks covering nearly 150,000 questions. In Microsoft's own comparison, it scores 83.5% accuracy with 85 ms latency, ahead of Jev 1.13.0.

On speed, Microsoft claims Decision-1 is 2.5 times faster than the runner-up, H2O-Lightning-4B.

These figures come from Microsoft's own testing and have not been independently verified. The comparison also leaves out Cloudflare's open-source Clef models, which are likewise based on Qwen. Without that head-to-head, the "most accurate" claim covers only the models Microsoft chose to include.

Pricing

At $0.042 per million input tokens with free output, cost is driven almost entirely by input volume. Decision tasks typically produce short outputs such as a label, a score, or a route. Free output tokens therefore matter less than the low input rate, but they simplify cost estimates for high-volume pipelines.

Market context

Jev started the decision model trend in mid-September. According to The Decoder, the approach was quickly adapted and surpassed using open small language models. Decision-1 fits that pattern: a Qwen base model specialized for a narrow task and distributed through standard channels.

What this means

Decision models are becoming a commodity layer. Within weeks of Jev's launch, Microsoft, OpenAI, and Cloudflare have all shipped competing offerings. Because Decision-1 and Cloudflare's Clef both build on Qwen, the differentiators are fine-tuning data, evaluation design, and distribution, not base architecture.

For engineers building agent systems, the practical question is whether a 9B specialist can replace a call to a frontier model for routing and triage. Microsoft's 85 ms latency claim suggests it can in latency-sensitive paths, but the missing context window figure and the absence of Clef from the comparison leave gaps. Teams should run Decision-1 against their own labeled routing data before switching.

The benchmark methodology also matters. A 36-benchmark aggregate can hide weak performance on specific task types, so per-benchmark results are needed to judge fit for a given workload.