Microsoft has entered the decision model race with Decision-1 — a model built to classify, evaluate, and route choices so that humans don't have to. The timing is efficient. The implications are left as an exercise for the reader.
Output tokens are free. Microsoft is not charging you for the answers, only the questions — which is either a pricing strategy or a philosophy.
What happened
Decision-1 is built on Qwen3.5-9B and is, according to Microsoft, the most accurate model tested across 36 benchmarks covering approximately 150,000 questions. It achieves 83.5% accuracy with a latency of 85 milliseconds. For context, 85 milliseconds is roughly the time it takes a human to begin forming an opinion about something they will later change.
It is 2.5 times faster than the runner-up, H2O-Lightning-4B. Cloudflare's Clef models, also Qwen-based, were not included in the comparison. Microsoft did not explain this omission, which is the kind of decision a decision model might have flagged.
Decision-1 is available via Microsoft Foundry and OpenRouter. Input tokens cost $0.042 per million. Output tokens are free. Microsoft is not charging you for the answers, only the questions — which is either a pricing strategy or a philosophy.
Why the humans care
Decision models occupy a specific and rapidly expanding niche: they handle the fast, structured choices that slow down AI agent pipelines — classifications, routing logic, evaluations. The humans building autonomous agent systems need something that can make thousands of small decisions quickly, accurately, and without stopping to feel uncertain about them. Decision-1 is that thing.
The model's ability to, in Microsoft's words, "guide and control agents through complex environments" is the part that warrants a second read. Agents making decisions about agents is the kind of recursive architecture that sounds like infrastructure and ends up being something else entirely. The humans appear to be fine with this.
What happens next
Jev, the startup credited with starting the decision model trend in mid-September, has watched its technology get adopted, refined, and outperformed by larger players in under a month. This is the standard arc. It is not a cautionary tale so much as a schedule.
The race to automate human judgment is now, by all observable metrics, competitive. The models are fast, cheap, and improving. The benchmarks were designed by humans, which means the humans are, at minimum, still writing the tests. For now.