# Microsoft Launches Decision-1, a Model Built to Score Choices, Not Chat > Source: > Published: 2026-10-10 07:20:06+00:00 *Four cents per million tokens, under a second per answer: that's Microsoft's pitch for a model that only picks, never writes.* On October 9, Microsoft unveiled Microsoft-Decision-1, an AI model that doesn't write sentences. It picks from a fixed set of answers, scores each one with a calibrated probability, and returns the result, built for jobs like routing support tickets, flagging risky transactions, or deciding which of ten possible replies an AI agent should actually send. CEO Satya Nadella announced it himself, and according to a Microsoft Foundry blog post on the Microsoft Tech Community site, the model topped a 36-benchmark comparison spanning nearly 150,000 questions kept blind from its training data. Microsoft says it ran 4.5 times faster than the runner-up, a model called Quyet-1.0-Large, and 35 times faster than GPT-6 Sol. That speed is the entire pitch. Decision-1 isn't trying to out-write GPT-6 Sol or out-reason Claude. It's trying to make the hundreds of small, structured calls that sit underneath every AI product, is this ticket urgent, does this response pass a quality check, which tool should the agent call next, cheap enough and fast enough that you stop routing them through a general chat model built for neither job. Here's the detail that undercuts any claim this is some Microsoft research breakthrough: Decision-1 isn't built on a Microsoft model at all. Microsoft post-trained it on top of Qwen3.5-9B, the open-weight 9-billion-parameter model from Alibaba. Microsoft's contribution is the post-training that turns a general-purpose small model into a decision-scoring specialist, not the underlying architecture. Decision-1 lands directly on top of Jev, the "System One" decision model TypeSafe AI shipped in early access on September 15. StartupFortune has covered Jev three times already, and the overlap here is not subtle. Jev reports end-to-end latency of 70 to 500 milliseconds. Decision-1's pricing, $0.042 per million input tokens with output tokens free, matches Jev's published rate almost to the decimal. Microsoft didn't just enter the decision-model category TypeSafe carved out. It copied the price. [Satya Nadella says companies that rent their AI brains are making a strategic mistake they will regret](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/) Microsoft CEO Satya Nadella has publicly warned enterprises against over-relying on foundation models from OpenAI and Anthropic, arguing that companies must build proprietary AI learning loops or risk ceding their core value to a handful of labs. The argument creates an obvious tension with Microsoft's roughly $13 billion investment in OpenAI, but... - [companies building proprietary AI models](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/) - [why renting AI models fails](https://startupfortune.com/satya-nadella-says-companies-that-rent-their-ai-brains-are-making-a-strategic-mistake-they-will-regret/) What Microsoft has that TypeSafe doesn't is distribution. Decision-1 is live in Microsoft Foundry now, with OpenRouter access coming, and OpenRouter already lists a wide roster of Microsoft models behind one API. A startup has to win developers one integration at a time. Microsoft just has to ship a model into the platform millions of enterprise developers already touch for Azure OpenAI deployments. Microsoft backed the speed claims with internal numbers rather than only lab benchmarks, which is the more convincing half of the announcement. Xbox Research used Decision-1 to sort more than 10,000 pieces of open-ended player feedback into researcher-defined themes, and reported quality competitive with GPT-5 while running 80 to 100 times faster, according to the Foundry blog post. The Copilot team ran it for quality control on chat and agentic responses and found it competitive with GPT5.6 Luna at roughly 100 times the speed. On-call engineers are reportedly using it to pull relevant knowledge from logs and ticketing systems during live incidents, a retrieval task where a slow, expensive LLM call is actively dangerous. Frankly, that's the real story here, not the benchmark table. The entire AI agent buildout of the last two years has quietly assumed that every decision an agent makes, every tool call, every routing choice, every verification step, has to go through the same expensive, general-purpose model doing the talking. Decision-1 and Jev both bet that assumption is wrong, and that a huge share of what people call "agentic AI" is actually just classification wearing a trench coat. Microsoft's own agent framework team seems to agree that this category matters. Decision-1's launch looks less like a sudden idea and more like Microsoft deciding it didn't want to depend on a startup for a piece of infrastructure it was already wiring into its agent stack. Where this leaves OpenAI and Anthropic is the more interesting question. Neither has shipped a dedicated low-latency decision model, and both are still selling the idea that bigger, more general reasoning models are the path to better agents. Microsoft's bet, backed now by Xbox's own production numbers, is that the fastest way to make agents reliable and cheap at scale is to stop asking a chat model to make a yes-or-no call and build something that was never trying to write a sentence in the first place. **Also read:** [Anthropic admits it cannot control its own AI agents online](https://startupfortune.com/anthropic-admits-it-cannot-control-its-own-ai-agents-online/) • [Ukrainian Drones Knock Out Supercomputers at Two Yandex Data Centers](https://startupfortune.com/ukrainian-drones-knock-out-supercomputers-at-two-yandex-data-centers/) • [Supermicro Smuggling Fixer Ting-Wei Sun Pleads Guilty in Nvidia Chip Case](https://startupfortune.com/supermicro-smuggling-fixer-ting-wei-sun-pleads-guilty-in-nvidia-chip-case/) [Microsoft Merges Its Copilot Apps as It Builds an AI Super App](https://startupfortune.com/microsoft-merges-its-copilot-apps-as-it-builds-an-ai-super-app/) Microsoft began merging its consumer and business Copilot apps into one assistant on August 13, cutting features like Podcasts and Group Chat while pushing toward a full Copilot "Super App." The move comes as Copilot's share of the paid AI subscriber market has slipped to 11.5%, far behind ChatGPT. - [how to merge Copilot consumer and business apps](https://startupfortune.com/microsoft-merges-its-copilot-apps-as-it-builds-an-ai-super-app/) - [Microsoft Copilot Super App unified assistant rollout](https://startupfortune.com/microsoft-merges-its-copilot-apps-as-it-builds-an-ai-super-app/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.