Microsoft has entered the decision-model race with Decision-1, a small model that does not write text at all — it scores a fixed set of options and returns a calibrated probability for each one. It is generally available now through Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with output free. Why it matters: agents make thousands of tiny routing and checking decisions, and Microsoft is pricing that control layer at less than half of OpenAI’s competing endpoint.
October 11, 2026
What happened
Microsoft announced Decision-1 on October 9. The model is post-trained from Alibaba’s open-weight Qwen3.5-9B, accepts text only, works within a 32,768-token context, and answers in JSON in a single pass. Instead of generating a reply, it takes a situation, a question, and a defined list of options — yes/no, multiple choice, a rating scale, a classification, or a rubric — and attaches a probability to each. Supported jobs include routing requests, grading AI responses, checking whether an answer is grounded in supplied evidence, and scoring actions an agent proposes to take. It can also abstain when the evidence is not there.
The numbers Microsoft is claiming
In Microsoft’s own comparison across 36 benchmarks and 147,137 questions, Decision-1 averaged 83.5% accuracy, ahead of OpenAI’s GPT-6 Luna Decisions endpoint at 79.4% in the same chart. Microsoft reports median latency of 85 milliseconds, and says the model changed its answer in only 1.3% of cases when requests were paraphrased or the option order was shuffled. Internally, Microsoft says its Xbox Research team sorted more than 10,000 feedback items 14 times faster and 200 times cheaper than with GPT-6 Sol, and that the Copilot team used it for quality checks at up to 100 times the speed of its previous setup.
Two honest caveats. Every figure above comes from Microsoft’s own testing — the model does not yet appear on the independent JevBench leaderboard — and at least one rival, H2O.ai, disputes how its latency was presented in Microsoft’s chart. Treat the lead as claimed, not settled.
What it cannot do
The model card is clear about the limits: no text generation, no chat, no translation, no images or audio, and no explanations for its scores. Microsoft also says it must not be the sole decision-maker on credit, employment, housing, healthcare, or legal rights — applications are expected to define thresholds, escalation paths, and human oversight. Microsoft says future versions will be rebuilt on its own MAI models and OpenAI technology rather than Qwen.
Why it matters
Decision models are quietly becoming the traffic controllers of the agent era: cheap, fast checks that decide which model answers, whether an action is safe to run, and when a human needs to step in. We have covered Cloudflare’s Clef, AWS’s Strands Decider, and OpenAI’s Decisions API doing the same job. Microsoft’s entry matters because of the price — $0.042 per million input tokens against OpenAI’s $0.10 — and the distribution: it sits inside Foundry, where enterprise agent builders already work. If the accuracy holds up under independent testing, a lot of agent plumbing gets cheaper this month. Everything AI, tested — and on this one, the independent test results are the part still missing.
Frequently asked questions
What is a decision model?
An AI model that scores a closed set of options instead of generating text. Given a question and fixed possible answers, it returns a probability for each, so software can act on the result immediately.
How much does Microsoft Decision-1 cost?
$0.042 per million input tokens, with output tokens free, through Microsoft Foundry and OpenRouter. OpenAI’s comparable Luna decisions endpoint charges $0.10 per million input tokens.
Is Decision-1 open source?
No. It is a hosted, proprietary API, although its base model, Alibaba’s Qwen3.5-9B, is open-weight. Several rivals, including Cloudflare’s Clef, are open source.
Can Decision-1 replace a chatbot model?
No. It does not generate text, hold conversations, or explain its reasoning. It is designed to sit alongside larger models and handle routing, classification, and verification.
Sources: Microsoft announcement and Foundry model card (via MarkTechPost and The Decoder), AI Weekly, TestingCatalog