Microsoft has introduced a model designed to make constrained choices rather than produce open-ended prose, joining a widening field of companies that see faster and cheaper decision systems as a building block for AI software. The Register reported on October 10 that Microsoft-Decision-1 is available through Microsoft Foundry and is expected to reach OpenRouter soon.

The distinction is important. Conventional large language models are typically asked to generate text or work through broad problems. Decision models instead take a defined set of possible answers and assign probabilities to them. That narrower format can suit classification, routing, prioritization and other software tasks in which an application needs a structured result it can act on immediately.

An AI foundation block feeds a decision layer that routes ranked probability tokens.
The first Microsoft-Decision-1 version is built on Qwen3.5-9B and designed for structured choices.

Microsoft built the first version by post-training Qwen3.5-9B, an open-weight model from Alibaba Cloud’s Qwen family, according to The Register. Microsoft has not explained why it chose that foundation. The company said it plans to rebase later versions on models from Microsoft and OpenAI, signaling that the underlying model may change even as the decision-oriented product remains.

The launch arrives amid an unusually crowded burst of activity. The Register listed decision products or models from OpenAI, Cloudflare, Strands, Liquid AI, Perplexity, Snowflake, Surogate and H2O.ai, among others. It said more than 100 models are already competing for attention. TypeSafe AI’s Jev, announced three weeks earlier, helped focus interest on systems that return a limited range of probability-rated responses.

Microsoft is positioning Decision-1 on speed, accuracy and cost, but the headline figures are the company’s own measurements. The Register said Microsoft claims the model was 2.5 times faster than H2O-Lightning-4B and 2.8 times faster than Jev in its latency test. Microsoft also reported 83.5 percent accuracy across 36 benchmarks and a 92.2 percent confidence score, second to Quyet-1.0-Large in that comparison.

Multiple decision engines race along data tracks beneath a calibration gauge.
Vendor-reported speed and accuracy will still need testing on real workloads and failure cases.

The pricing is intended to make the model attractive for high-volume classification. According to the report, Microsoft charges $0.042 per million input tokens and nothing for output tokens. The company further claims Decision-1 is more than 20 times cheaper than OpenAI’s GPT-6 Sol for text-classification work. Those comparisons should be read as vendor-reported results until buyers test the model against their own data and operating conditions.

The narrower output does not eliminate risk. The Register noted that decision models can still make errors, even if advocates argue they avoid the free-form hallucinations associated with generative systems. A confident probability score can be useful to software, but it can also make a wrong classification easier to automate at scale. Accuracy, calibration and the cost of mistakes will therefore matter as much as raw latency.

Microsoft’s move shows how quickly AI vendors are dividing workloads between general-purpose generators and specialized systems. If decision models prove dependable, they could become inexpensive control layers inside larger agent workflows, sending requests to the right service, checking proposed actions or ranking possible next steps. For now, Decision-1 adds a major platform provider to the race while leaving its performance claims to be tested beyond Microsoft’s benchmarks.