--- title: Microsoft-Decision-1: a cheap, text-only decision model... description: Microsoft has launched Microsoft-Decision-1, a text-only decision-scoring model in Foundry built by post-training Alibaba's Qwen3.5-9B, arriving three days... canonical: https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Microsoft-Decision-1: a cheap, text-only decision model built on Qwen | daily.dev og:description: Microsoft has launched Microsoft-Decision-1, a text-only decision-scoring model in Foundry built by post-training Alibaba's Qwen3.5-9B, arriving three days... og:url: https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7 og:image: https://api.daily.dev/og/posts/UQZhBc7P7.png og:image:alt: Microsoft-Decision-1: a cheap, text-only decision model built on Qwen og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Microsoft-Decision-1: a cheap, text-only decision model built on Qwen **[Collections](https://daily.dev/sources/collections)** · 6 min read · 0 upvotes · 0 comments ## Summary Microsoft has launched Microsoft-Decision-1, a text-only decision-scoring model in Foundry built by post-training Alibaba's Qwen3.5-9B, arriving three days after OpenAI opened its Decisions API to all developers. It costs $0.042 per million input tokens with no output charge, matching TypeSafe's Jev pricing, and is already used internally in Xbox Research, Copilot, on-call incident response, and Microsoft Discovery. It lacks image analysis (unlike OpenAI and Cloudflare offerings), has no announced open weights, and its System One API compatibility is unconfirmed. Adversarial testing showed only 1.3% of answers changed under perturbation, though it hasn't faced the tougher JevOut methodology that flipped most of a rival model's decisions. The piece closes by questioning accountability as cheap automated decisions scale across more applications. ## Content Microsoft has released Microsoft-Decision-1, a small model that doesn't write text. You give it a state and a fixed set of allowed answers, and it returns probabilities over those answers. It is meant for routing, classification, prioritization, verification, rubric grading, and the "retry or escalate?" step in an agent loop. It won't write the implementation, explain itself, or replace the model carrying the conversation. Microsoft built it by post-training Alibaba's open-weight Qwen3.5-9B. It's available through Microsoft Foundry and OpenRouter, and Microsoft says it will later rebase the model on OpenAI and its own MAI models. That is a roadmap, not something you can use today. ## Where it sits in a crowded field The launch came three days after OpenAI opened its Decisions API to all developers. The same day, TypeSafe announced an $870 million Series A at a $7.5 billion valuation. TypeSafe makes Jev, the closest rival. Cloudflare, Strands, Liquid AI, Perplexity and Snowflake have offerings too. One gap: Decision-1 is text-only. OpenAI's API and Cloudflare's Clef-omni handle images (Clef-omni also takes audio and video), so checking product photos for damage isn't possible here. No open weights have been announced, and Microsoft hasn't confirmed compatibility with the emerging System One API standard. ## Price Both Decision-1 and Jev list $0.042 per million input tokens, with free output. At 1,000 billed input tokens per decision, 100,000 decisions cost $4.20. Per-token price isn't per-decision price, though. In one independent billing log, the same ticket question was 75 input tokens on Microsoft and 360 on Jev. Across 700 normal-condition classification requests, mean billed input was 127 tokens for Microsoft and 431 for Jev. That works out to about $0.0037 versus $0.0127 in total, or roughly 3.4x. The authors don't know why the counts differ, and the figure shouldn't be read as a universal saving. Measure your own prompts. The bigger costs usually come from a wrong cheap route, a needless escalation, or a bad retry. Context windows differ too: OpenRouter lists 32,768 tokens for Microsoft and 64,000 for Jev 1.13. ## Microsoft's own benchmark Microsoft compared 36 benchmarks covering 147,137 questions, which it says were kept blind from training. It claims first place overall. - Average accuracy: Decision-1 83.5%, Jev 82.3% - Calibration score (higher is better): Decision-1 92.2, Jev 93.7 - Adjusted P50 latency: Decision-1 85 ms, Jev 240 ms That's a 1.2-point accuracy lead, and Jev wins on calibration. Microsoft also says the model runs 2.5x faster than runner-up H2O-Lightning-4B and 14x to 35x faster than GPT-6 Sol, depending on the task. The latency numbers mix serving conditions: Microsoft measured its own model on Foundry against JevBench measurements for the others. Live OpenRouter pages currently show about 0.26 s P50 for Microsoft's route and 0.16 to 0.17 s for Jev's. Those aren't controlled either, but they show that "faster" needs a route and a workload attached. Microsoft is already using the model internally: Xbox Research classified over 10,000 feedback items, and Copilot, on-call incident response and Microsoft Discovery are testing it. Those are Microsoft's claims, not independent trials. ## An independent test SOTAAZ Lab ran 500 TREC questions plus a 200-item AG News set through OpenRouter and published its logs and scripts. The twist was making labels contradict their definitions, for example an option called "location" that describes a person. | Condition | Decision-1 | Jev 1.13 | |---|---|---| | Normal labels | 482/500 (96.4%) | 467/500 (93.4%) | | Contradictory labels | 476/500 (95.2%) | 428/500 (85.6%) | With a 0.9 top-probability gate on contradictory labels, Decision-1 accepted 400 answers with 5 wrong. Jev accepted 213, also with 5 wrong. So Decision-1 handled semantic conflict better in this test. That doesn't establish general calibration or production safety. TREC is a short, clean six-class task, and Microsoft hasn't disclosed its training data, so prior exposure can't be ruled out. The archive also can't fully reproduce fresh inference, because its loader references missing files. Don't use the run as a speed test either: the authors paced their requests, and all 36 rate-limited attempts were on Microsoft's route, though every one eventually completed. Repeat calls weren't identical. Across 500 repeated Microsoft requests, one answer changed and probabilities shifted on 439, mostly slightly. Log the returned model version and re-test after updates. On the Jev side, Tobias Deußer, Lorenz Sparrenberg and Rafet Sifa evaluated it across 37 datasets and 346,009 requests. Tuning binary thresholds lifted UNFAIR-ToS micro-F1 from 0.50 to 0.75. Neither this nor a separate routing experiment says anything direct about Decision-1, but both argue for tuning thresholds and checking what the decision model actually contributes. A caution on thresholds: Jev's "confidence" is not its top probability. TypeSafe defines it as (p_max − 1/n) / (1 − 1/n), so with three options a top probability of 0.9 becomes 0.85. SOTAAZ gated on top probability. Don't carry a threshold across fields, providers or tasks without validating it. ## Adversarial robustness In adversarial testing, Decision-1's answers changed only 1.3% under perturbation. It hasn't faced the more aggressive JevOut methodology, which flipped the majority of a rival model's correct decisions. ## Integrating it This is a typed service, not a chat endpoint. A request has a model, a state to judge, and named questions: Choice (options mapped to descriptions), Score (ordered rubric levels), or Noul (is this statement true?). On OpenRouter the model is `microsoft/microsoft-decision-1`. Foundry uses `/providers/microsoft/v1/systemone`, with your deployment name in the model field and Entra auth or an API key. Swapping a chat model name won't work. Keep questions atomic. "Should we ship?" mixes correctness, authorization and risk. Ask narrow questions and combine them in ordinary code. Deterministic code, not the model, should handle identity, permissions and allowed side effects. For Codex, the practical route is a custom MCP adapter exposing a read-only advisory tool such as `score_candidate_action`. It calls the API, validates the answer against a fixed schema, and returns the distribution, chosen option, model version and billed usage. Register it with `codex mcp add decision-scorer -- `, then check `codex mcp list` and `/mcp`. This is your adapter, not an official Microsoft MCP server. Keep the credential in the environment, not in the TOML file. Hosted ChatGPT doesn't read local Codex config, and an MCP tool isn't guaranteed to run before every action, so it is advice, not a security boundary. Wherever you integrate, limit retries, log decisions, and fail closed on malformed responses or network errors. Include an explicit abstention option, since an out-of-scope request can still resemble a familiar answer. ## Before swapping out Jev Run a few hundred labeled decisions from your real workload, split into development and held-out sets. Compare Decision-1, Jev and a non-model baseline such as rules or a best-fixed-model router. Record per-class errors, dangerous false accepts, billed tokens, P50/P95 latency and rate limits, and test reordered options, misleading labels and prompt injection. Set your maximum false-accept rate and cost target in advance, then start in shadow mode. My read: it's a cheap, bounded scorer with encouraging robustness numbers, not a universal upgrade over Jev. It's also notable that Microsoft chose an Alibaba base model rather than an OpenAI one. ## Questions this post answers ### What is Microsoft Decision-1 and how does it price compared to other decision models? Microsoft Decision-1 is a text-only decision-scoring model launched in Microsoft Foundry, built by post-training Alibaba's Qwen3.5-9B rather than an OpenAI base model. It costs $0.042 per million input tokens with no charge for output, matching TypeSafe's Jev pricing. It scores predefined options for tasks like spam detection or routing, rather than generating open-ended text. _Developers comparing cheap decision-scoring APIs can track new model launches like this one on daily.dev._ ### Can Microsoft Decision-1 analyze images for decision-making tasks? No, Microsoft Decision-1 is text-only, unlike competing decision models from OpenAI and Cloudflare that can also analyze images, for example checking product photos for damage. Microsoft hasn't announced open weights for the model, and whether it supports the emerging System One API standard remains unconfirmed. _Anyone evaluating decision-model capabilities before integrating one can follow these gaps on daily.dev._ ### How robust is Microsoft Decision-1 against adversarial inputs? In adversarial testing, Decision-1's answers changed only 1.3% of the time under perturbation, a strong result on paper. However, it has not yet been tested against the more aggressive JevOut methodology, which flipped the majority of a rival model's correct decisions, so the 1.3% figure should be treated as a best-case number until that stress test runs. _Teams weighing reliability claims before trusting a decision model in production can follow robustness testing on daily.dev._ ## Similar posts on daily.dev - [Microsoft separates decision-making from generative AI with a new model](https://daily.dev/posts/microsoft-separates-decision-making-from-generative-ai-with-a-new-model-z7lrvkhp8) · InfoWorld · 0 upvotes · 0 comments - [Decision AI: The next level in agentic system architecture](https://daily.dev/posts/decision-ai-the-next-level-in-agentic-system-architecture-rvqpmyltg) · Medium · 2 upvotes · 0 comments - [Jev Doesn't Write Text. It Returns Probabilities.](https://daily.dev/posts/jev-doesn-t-write-text-it-returns-probabilities--3tcrb0ghc) · DEV · 7 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#microsoft](https://daily.dev/tags/microsoft), [#qwen](https://daily.dev/tags/qwen) [View this post on daily.dev](https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Microsoft-Decision-1: a cheap, text-only decision model built on Qwen","url":"https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7"},"datePublished":"2026-10-10T15:12:58.397Z","dateModified":"2026-10-14T01:32:25.870Z","description":"Microsoft has launched Microsoft-Decision-1, a text-only decision-scoring model in Foundry built by post-training Alibaba's Qwen3.5-9B, arriving three days...","image":"https://i.ytimg.com/vi/LB7RJ0ruFio/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/LB7RJ0ruFio/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,microsoft,qwen","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Microsoft-Decision-1: a cheap, text-only decision model built on Qwen"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/microsoft-decision-1-a-cheap-text-only-decision-model-built-on-qwen-uqzhbc7p7#faq","mainEntity":[{"@type":"Question","name":"What is Microsoft Decision-1 and how does it price compared to other decision models?","acceptedAnswer":{"@type":"Answer","text":"Microsoft Decision-1 is a text-only decision-scoring model launched in Microsoft Foundry, built by post-training Alibaba's Qwen3.5-9B rather than an OpenAI base model. It costs $0.042 per million input tokens with no charge for output, matching TypeSafe's Jev pricing. It scores predefined options for tasks like spam detection or routing, rather than generating open-ended text. Developers comparing cheap decision-scoring APIs can track new model launches like this one on daily.dev."}},{"@type":"Question","name":"Can Microsoft Decision-1 analyze images for decision-making tasks?","acceptedAnswer":{"@type":"Answer","text":"No, Microsoft Decision-1 is text-only, unlike competing decision models from OpenAI and Cloudflare that can also analyze images, for example checking product photos for damage. Microsoft hasn't announced open weights for the model, and whether it supports the emerging System One API standard remains unconfirmed. Anyone evaluating decision-model capabilities before integrating one can follow these gaps on daily.dev."}},{"@type":"Question","name":"How robust is Microsoft Decision-1 against adversarial inputs?","acceptedAnswer":{"@type":"Answer","text":"In adversarial testing, Decision-1's answers changed only 1.3% of the time under perturbation, a strong result on paper. However, it has not yet been tested against the more aggressive JevOut methodology, which flipped the majority of a rival model's correct decisions, so the 1.3% figure should be treated as a best-case number until that stress test runs. Teams weighing reliability claims before trusting a decision model in production can follow robustness testing on daily.dev."}}]} ```