{"slug": "microsoft-decision-1-score-agent-choices-not-chat", "title": "Microsoft Decision-1: Score Agent Choices, Not Chat", "summary": "Microsoft shipped Decision-1 on October 9, a 9B-parameter model in Microsoft Foundry that scores a fixed set of choices and returns calibrated probabilities as JSON instead of generating text, priced at $0.042 per million input tokens with no output tokens. Microsoft says Decision-1 runs roughly 35x faster than GPT-6 Sol at median latency and is about 70x cheaper per call, but the company has published no third-party benchmarks or reproducible methodology. Decision-1 is post-trained on Alibaba's open-weight Qwen3.5-9B, and Microsoft says it will \"soon rebase on models including Microsoft AI and OpenAI,\" a change that would shift model behavior for anyone integrating it today.", "body_md": "Microsoft shipped a model on October 9 that never says anything. [Microsoft-Decision-1](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-microsoft-decision-1-in-microsoft-foundry-for-decision-and-classific/4562742) doesn’t generate text, doesn’t explain its reasoning, doesn’t chat. You hand it a state and a set of choices; it returns probabilities. That’s the entire product. For developers who’ve been routing agent pipelines through frontier LLMs because nothing cheaper existed, this is the architecture change worth paying attention to.\n\n## What Decision-1 Actually Does\n\nDecision-1 exposes a structured Decisions API — not a chat endpoint. You submit a context (the “state”), a question, and a fixed choice set. The model scores each option and returns calibrated probabilities as JSON. Zero output tokens. There’s nothing to parse, no regex to write, no prompt template to maintain.\n\nThree question types are supported: `noul` for yes/no probability, `choice` for picking one option from a set, and `score` for a value on an ordered scale. Here’s what a safety gate looks like:\n\n```\ncurl \"$AZURE_ENDPOINT/providers/microsoft/v1/systemone\" \\\n  -H \"Authorization: Bearer $TOKEN\" \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\n    \"model\": \"my-decision-1-deployment\",\n    \"state\": \"User requests deletion of all production records.\",\n    \"questions\": {\n      \"safe\": { \"type\": \"noul\", \"question\": \"Should this proceed?\" }\n    }\n  }'\n```\n\nThis is architecturally different from prompting a small LLM and parsing its reply. Decision-1 can’t hallucinate its way into a wrong branch — there’s no generation path to go off-rails.\n\n## The Cost Makes It a Background Primitive\n\nPricing: $0.042 per million input tokens. Output is free because there is none. At that price, running Decision-1 on every agent branch call adds up to almost nothing. A pipeline making 10,000 routing decisions per hour spends under $0.05. Compare to GPT-6 Sol at around $3 per million input tokens — that’s roughly 70x cheaper per call.\n\nThat price point changes how you design. When routing costs essentially nothing, you can add decision gates everywhere in your pipeline without a cost conversation: before every tool call, after every output, at every handoff between agents. That’s the value proposition.\n\n## They Built It on Alibaba’s Model\n\nThis is the detail Microsoft would prefer you to skim past. Decision-1 is post-trained on Alibaba’s open-weight **Qwen3.5-9B**, not on Microsoft’s own MAI models or anything from its OpenAI partnership. Microsoft hasn’t explained the choice publicly. The company does say it will “soon rebase on models including Microsoft AI and OpenAI.”\n\nThe practical concern isn’t geopolitical — it’s versioning. When Microsoft rebases, the model’s behavior changes. Anyone integrating Decision-1 today is buying a current behavior that will silently shift in a future version. Pin to the version identifier in your deployment config, and plan to re-validate accuracy whenever Microsoft ships a new base.\n\n## Speed: 35x Faster, No Benchmarks to Verify\n\nMicrosoft claims Decision-1 runs approximately 35x faster than GPT-6 Sol at median latency. That’s credible on its face — single-pass scoring over a fixed choice set is orders of magnitude lighter than autoregressive generation. But [Microsoft has not published third-party benchmarks](https://www.orcarouter.ai/blog/microsoft-decision-1-ga-foundry-no-published-benchmarks) or reproducible methodology. Take the number as directionally true until you measure it yourself against your actual workload.\n\n## Where It Fits (and Doesn’t)\n\nDecision models are now a category with several options. The quick comparison:\n\n| Model | Size | Context | Weights | Price/M in | \n|---|---|---|---|---|\n| Clef (Cloudflare) | 27B | 64K | Apache 2.0 | Workers AI pricing | \n| Strands Decider 2B (Amazon) | 2B | 16K | Apache 2.0 | Free (open weights) | \n| Microsoft Decision-1 | 9B | 32K | Closed | $0.042 | \n| OpenAI Decisions API | Unknown | Unknown | Closed | Not published | \n\nDecision-1’s real advantage is Azure integration. If you’re already on Foundry, it drops in with your existing auth, billing, and RBAC — no new vendor to manage. Its disadvantages versus [Cloudflare Clef](https://thetechlantis.com/news/cloudflare-clef-open-decision-models): no multimodal support, smaller context window (32K vs 64K), closed weights. For offline or CPU-constrained environments, Amazon’s Strands Decider 2B is fully open and considerably smaller.\n\nOn OpenRouter, Decision-1 ranks 6th of 14 decision models — solid but not dominant. Worth knowing before you commit it to production as your default.\n\n## Getting Started\n\nDecision-1 is available now in [Microsoft Foundry](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision). Deploy from the model catalog, copy your resource endpoint and deployment name, then POST to `/providers/microsoft/v1/systemone` with a Microsoft Entra ID token or API key. OpenRouter support is listed as coming soon.\n\nIf you’re not on Azure, Clef and Strands Decider are open-weight and immediately runnable. The decision model category is healthy enough that you’re not locked into any single vendor — which is exactly how this layer of your agent stack should work.", "url": "https://wpnews.pro/news/microsoft-decision-1-score-agent-choices-not-chat", "canonical_source": "https://byteiota.com/microsoft-decision-1-agent-routing/", "published_at": "2026-10-12 16:08:07+00:00", "updated_at": "2026-10-12 16:17:40.290336+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-products", "ai-infrastructure"], "entities": ["Microsoft", "Microsoft Decision-1", "Microsoft Foundry", "Alibaba", "Qwen3.5-9B", "GPT-6 Sol", "OpenAI", "Cloudflare"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/microsoft-decision-1-score-agent-choices-not-chat", "markdown": "https://wpnews.pro/news/microsoft-decision-1-score-agent-choices-not-chat.md", "text": "https://wpnews.pro/news/microsoft-decision-1-score-agent-choices-not-chat.txt", "jsonld": "https://wpnews.pro/news/microsoft-decision-1-score-agent-choices-not-chat.jsonld"}}