{"schema":"sanctity-signal/0.2","record_type":"signal","id":"microsoft-decision1-qwen-e047","domain":"models","drop_id":"signal-e047","occurred_at":"2026-10-09","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Microsoft","action_type":"release of a first-party branded decision model post-trained from Alibaba's Qwen3.5-9B open weights, priced at the same rate per million input tokens as a 24 day old startup's model","observed":"Microsoft published Introducing Microsoft-Decision-1 in Microsoft Foundry for decision and classification workloads on October 9, 2026, on commandline.microsoft.com with a published time of 2026-10-10 01:35 UTC, and on the Azure AI Foundry blog, modified October 9 at 14:05 Pacific time. The author is Achint Srivastava, vice president of software engineering in Microsoft's Office of the CTO. On provenance: \"we post trained Qwen3.5-9B for fast, single-pass decision scoring.\" Qwen is developed by Alibaba Cloud. On price: \"Input tokens cost $0.042 USD per million tokens. Output tokens are free.\" Microsoft says that is more than 20 times cheaper than GPT-6 Sol on text classification. On latency: \"Microsoft-Decision-1 P50 latency is ~35x faster than GPT-6 Sol,\" a reported median of 85 milliseconds against 3.01 seconds, and 4.5 times faster than the runner up, Quyet-1.0-Large. On accuracy, Microsoft reports leading a 36 benchmark comparison spanning nearly 150,000 withheld questions at 83.5%, against 81.9% for Quyet-1.0-Large, 79.4% for GPT-6 Luna Decisions and 77.2% for H2O-Lightning-4B, and second on calibration at 92.2. On stability: \"Microsoft-Decision-1 changes its decision on 1.3% of perturbations on average,\" with no flips when options were paraphrased, reversed or shuffled. On safety: \"We tested Microsoft-Decision-1 on 5,250 requests across 11 benchmarks,\" covering harmful content, jailbreaks and prompt injection, with no score published. Internal results Microsoft reports: Xbox Research labelled over 10,000 feedback items at quality competitive with GPT-6 Sol, over 14 times faster and 200 times cheaper; the Copilot team found it competitive with GPT-5.6 Luna and 100 times faster; Microsoft Discovery found scoring 46 times more consistent than an LLM approach. It is generally available in Foundry as a direct from Azure model, pay as you go, with OpenRouter coming. It returns JSON, has a 32,768 token context, publishes no weights, and its model card bars sole automated decisions on credit, employment, housing, healthcare or legal rights. Microsoft says it \"will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.\" Not established: the reason for leaving Qwen, the hardware, the training data mix beyond public and synthetic datasets, and any independent check of the figures. H2O disputes the latency method, saying its measured median is 29 milliseconds rather than 210.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Microsoft: Introducing Microsoft-Decision-1 in Microsoft Foundry, Achint Srivastava, Oct 9, 2026","url":"https://commandline.microsoft.com/microsoft-decision-1-model-foundry/"},{"role":"primary","label":"Microsoft Azure AI Foundry blog: the same announcement, modified Oct 9, 2026","url":"https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-microsoft-decision-1-in-microsoft-foundry-for-decision-and-classific/4562742"},{"role":"corrob","label":"The Register: Microsoft leans on open weight model from Chinese AI lab to challenge Jev, Oct 10, 2026","url":"https://www.theregister.com/ai-and-ml/2026/10/10/microsoft-leans-on-open-weight-model-from-chinese-ai-lab-to-challenge-jev/5302473"},{"role":"corrob","label":"MarkTechPost: Microsoft releases Microsoft-Decision-1, a Qwen3.5-9B decision model, Oct 9, 2026","url":"https://www.marktechpost.com/2026/10/09/microsoft-ai-releases-microsoft-decision-1-a-qwen3-5-9b-decision-scoring-model/"},{"role":"corrob","label":"Liquid AI: Open d1, edge decision models for text, vision and audio","url":"https://www.liquid.ai/blog/d1-open"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Low","persistence":"Year","rationale":"Cost is medium: post-training a 9 billion parameter open model is cheap next to frontier training, but free output tokens on a generally available Azure service are a continuing giveaway, and publishing a 36 benchmark comparison invites the scrutiny H2O has already given it. Commitment is high: generally available, priced, branded with the company name and signed by a named vice president, not a preview. Reversibility is low in the way that matters: the price is public and competitors now hold it, and a published price is harder to raise than a preview is to cancel. Persistence is a year, by Microsoft's own statement that it will rebase the model."},"shift":{"prior":"Microsoft's own-brand models run on Microsoft or OpenAI technology.","action":"Microsoft published Introducing Microsoft-Decision-1 in Microsoft Foundry for decision and classification workloads on October 9, 2026, on commandline.microsoft.com with a published time of 2026-10-10 01:35 UTC, and on the Azure AI Foundry blog, modified October 9 at 14:05 Pacific time. The author is Achint Srivastava, vice president of software engineering in Microsoft's Office of the CTO. On provenance: \"we post trained Qwen3.5-9B for fast, single-pass decision scoring.\" Qwen is developed by Alibaba Cloud. On price: \"Input tokens cost $0.042 USD per million tokens. Output tokens are free.\" Microsoft says that is more than 20 times cheaper than GPT-6 Sol on text classification. On latency: \"Microsoft-Decision-1 P50 latency is ~35x faster than GPT-6 Sol,\" a reported median of 85 milliseconds against 3.01 seconds, and 4.5 times faster than the runner up, Quyet-1.0-Large. On accuracy, Microsoft reports leading a 36 benchmark comparison spanning nearly 150,000 withheld questions at 83.5%, against 81.9% for Quyet-1.0-Large, 79.4% for GPT-6 Luna Decisions and 77.2% for H2O-Lightning-4B, and second on calibration at 92.2. On stability: \"Microsoft-Decision-1 changes its decision on 1.3% of perturbations on average,\" with no flips when options were paraphrased, reversed or shuffled. On safety: \"We tested Microsoft-Decision-1 on 5,250 requests across 11 benchmarks,\" covering harmful content, jailbreaks and prompt injection, with no score published. Internal results Microsoft reports: Xbox Research labelled over 10,000 feedback items at quality competitive with GPT-6 Sol, over 14 times faster and 200 times cheaper; the Copilot team found it competitive with GPT-5.6 Luna and 100 times faster; Microsoft Discovery found scoring 46 times more consistent than an LLM approach. It is generally available in Foundry as a direct from Azure model, pay as you go, with OpenRouter coming. It returns JSON, has a 32,768 token context, publishes no weights, and its model card bars sole automated decisions on credit, employment, housing, healthcare or legal rights. Microsoft says it \"will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.\" Not established: the reason for leaving Qwen, the hardware, the training data mix beyond public and synthetic datasets, and any independent check of the figures. H2O disputes the latency method, saying its measured median is 29 milliseconds rather than 210.","update":"A Microsoft-branded model can ship on a Chinese laboratory's open weights."},"inference":{"established":["Microsoft post-trained Qwen3.5-9B for single-pass decision scoring.","Input costs $0.042 per million tokens, and output tokens are free.","Microsoft says it will rebase the model on Microsoft AI and OpenAI models."],"unknown":["That the accuracy claims hold. The 83.5% figure is Microsoft's own.","Why it leaves Qwen. Microsoft disclosed no reason for the move.","That TypeSafe loses. Neither company has published revenue or customers."]},"narrative":{"hook":"Microsoft shipped a first-party AI model built on Alibaba's Qwen, at a startup's exact price.","plain":"Microsoft-Decision-1 is post-trained from Qwen3.5-9B and costs $0.042 per million input tokens, output free. That is the price TypeSafe AI set for Jev on September 15.","story":["Microsoft named it Microsoft-Decision-1 and says it will rebase on its own and OpenAI models later. For now the brand sits on Alibaba's weights, and the price matches a 24 day old startup to three decimal places. Price, not capability, is the field this one is being fought on."],"trail":[{"d":"SEP 15","t":"TypeSafe AI launches Jev on a reported $40 million seed. It answers only with a yes or no, a choice from a list or a score, each with a confidence number, at $0.042 per million input tokens with output free. Edition 046."},{"d":"SEP 25","t":"Meituan releases a Chinese model built to run inside American developer tools, an early sign that Western products would carry Chinese weights. Edition 033."},{"d":"OCT 7","t":"Liquid AI publishes open weight decision models, d1-3B and the multimodal d1-omni-600M, on Hugging Face. A third actor in the same category inside 3 weeks."},{"d":"OCT 9","t":"TypeSafe AI is priced at $7.5 billion on $870 million led by Andreessen Horowitz, 24 days after launching Jev. Edition 046's capital card."},{"d":"OCT 9","t":"The same day, Microsoft publishes Microsoft-Decision-1, post-trained from Alibaba's Qwen3.5-9B, at $0.042 per million input tokens with free output, generally available in Foundry."},{"d":"Reacts","r":true,"t":"TypeSafe AI, whose price is now its competitor's price; Andreessen Horowitz and Sequoia Capital, who funded that price; Alibaba Cloud, whose weights now carry a Microsoft nameplate; OpenAI, whose Decisions API is in beta and whose partner skipped it; H2O, which disputes the latency method; and officials arguing about Chinese models in American systems."}],"glossary":[{"term":"Decision model","def":"A model that scores a fixed set of options and returns a probability for each, instead of writing text. Cheap to run, and easy to wire into software that needs an answer, not prose."},{"term":"Post-training","def":"Further training of an existing model to specialise it. The base model's origin stays in the product, which is why the Qwen lineage matters here."},{"term":"Open weights","def":"Model parameters anyone may download and build on, under a licence. They let one company ship a product on another company's research."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Decision models are a real category","text":"3 actors shipped one in 25 days, and prices are converging on a floor.","abstain":false},{"id":"b","label":"A cheap feature, not a category","text":"A 9 billion parameter classifier is a small, ordinary piece of software.","abstain":false},{"id":"c","label":"A deliberate squeeze on a startup","text":"Matching $0.042 exactly reads as a price aimed at one competitor.","abstain":false},{"id":"d","label":"Normal platform commoditisation","text":"Clouds always build the cheap tier of whatever their customers buy.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"c","direction":"strengthens","observable_event":"TypeSafe AI cuts Jev's input price below $0.042, or a fourth vendor lists a decision model at or under $0.042 per million input tokens, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"a","direction":"strengthens","observable_event":"Microsoft publishes usage or revenue for Microsoft-Decision-1, or OpenAI moves its Decisions API out of public beta with a listed price, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"Microsoft withdraws Microsoft-Decision-1, folds it into an existing model family without a separate price, or raises the input price, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"}]}