What is Microsoft-Decision-1?
Microsoft-Decision-1 is Microsoft's first decision model, released on 9 October 2026. Instead of writing text it takes a state, text or JSON, and one or more typed questions, and returns numbers: a probability for a yes-or-no question (Microsoft calls the type "noul"), a chosen option with a probability for every option, or a position on an ordered scale. It is a post-trained version of Alibaba's open Qwen3.5-9B.
In Foundry it is called through a plain HTTP endpoint at /providers/microsoft/v1/systemone, with Entra ID or an API key, on GlobalStandard or DataZoneStandard deployments. It is also on OpenRouter. Pricing is $0.042 per million input tokens with output free, the same as TypeSafe's Jev.
Microsoft says it was the most accurate of the models it tested across 36 benchmarks and nearly 150,000 held-out questions, at 83.5% average accuracy and a median response time of 85 milliseconds. Those figures are Microsoft's own.
Is Microsoft-Decision-1 worth it?
Microsoft-Decision-1 matters because of where it sits. Decision models, which return a typed choice or probability instead of prose, have mostly come from start-ups; this one is in the Foundry catalogue next to GPT-6 and Claude, under Azure's identity, networking and data terms, which is what many enterprise buyers need before they will use one.
Microsoft's numbers are strong. Across 36 benchmarks and 147,137 held-out questions it reports 83.5% average accuracy, ahead of Jev 1.13.0 at 82.3% and GPT-6 Luna Decisions at 79.4%, with an 85-millisecond median response that Microsoft calls 2.5 times faster than the runner-up, H2O-Lightning-4B. Microsoft measured its own latency inside Foundry; eesel's reading of OpenRouter's figures puts the median nearer 0.19 to 0.27 seconds. Microsoft has not published which benchmarks it used or how rivals were configured, and The Decoder notes that Cloudflare's open Clef models were left out of the comparison.
On price it ties Jev at $0.042 per million input tokens and costs about twice Perplexity's pplx-decider-v1.2 at $0.02, which also ships open weights. Decision-1's weights are not released, even though its base, Qwen3.5-9B, is open. A billion input tokens a month costs $42.
The request format, a state plus typed questions sent to a "systemone" endpoint, differs from OpenAI's Decisions API, which Foundry has sold on GPT-6 Luna since 22 October at $0.10 per million input tokens. Moving between the two means rewriting requests. Decision-1 takes text and JSON only, so image decisions need Luna.
Microsoft's documentation is candid about the limits: scores shift with how questions and options are worded or ordered, poorly framed questions still return confident numbers, and calibration is strongest on familiar task types. It advises human review for decisions about credit, employment, housing, healthcare or legal matters.
Where it earns its place
A support operations team at a SaaS company on Azure
Route incoming tickets to billing, engineering or account teams
Send each ticket as state with a choice question; escalate to a person when the top probability is low. At $0.042 per million input tokens, 2 million tickets of 500 tokens cost about $42 a month.
A platform team running coding and data agents
Block agent actions that touch production data
Ask a yes/no question about each proposed tool call before it runs and set a probability threshold from labelled examples, as Microsoft's docs advise; budget for 200 to 300 ms per check outside Foundry's own region, according to eesel's reading of OpenRouter's latency figures.
An evaluation lead grading model outputs
Score thousands of answers against a rubric every night
Use a score question with ordered levels such as Poor to Excellent; Microsoft recommends using scores for relative ranking and thresholds rather than as absolute ratings.
Before you commit
- Setup
- Deploy Microsoft-Decision-1 from the Foundry model catalogue, or with az cognitiveservices account deployment create, then POST state and questions to /providers/microsoft/v1/systemone; no model SDK is needed.
- Integrations
- Plain HTTP API on Foundry with Entra ID or API key; OpenRouter; Azure CLI deployment.
- Data and privacy
- Sold as a Foundry model by Azure, under which prompts are not used to train foundation models; DataZoneStandard keeps inference within a data zone, while GlobalStandard can process requests in any supported region.
- Cost at scale
- Input only, at $0.042 per million tokens: 10 billion tokens a month cost $420.
- Limits
- 32,768-token context; text or JSON state only; quotas per deployment with 429 errors when exceeded; scores sensitive to wording and option order; OpenRouter access runs through an alpha endpoint.
What’s good
- $0.042 per million input tokens, output free
- Reported 83.5% accuracy and 85 ms median latency across 36 benchmarks
- Foundry deployment with Entra ID and DataZoneStandard data residency
- Three typed question kinds: yes/no, choice and score, each with probabilities
What’s not
- Benchmarks are Microsoft's own and not reproducible
- Closed weights, though its Qwen3.5-9B base is open
- Text and JSON only; no image input
- Request format differs from OpenAI's Decisions API in the same catalogue
Best for
Azure teams that need cheap, fast classification, routing or gating decisions under Microsoft's data terms, and will validate thresholds on their own labelled data.
Microsoft-Decision-1 alternatives
Microsoft-Decision-1: common questions
- Is Microsoft-Decision-1 worth it?
- Microsoft-Decision-1 makes decision models a default Azure option at Jev's price, but its benchmark lead is self-reported and open rivals do the same job for less. We rate it 3.9 out of 5.
- How much does Microsoft-Decision-1 cost?
- $0.042 per 1M input tokens; output tokens free. Billed through an Azure subscription on Foundry, or through OpenRouter.
- What are the downsides of Microsoft-Decision-1?
- Benchmarks are Microsoft's own and not reproducible; closed weights, though its Qwen3.5-9B base is open; text and JSON only; no image input; request format differs from OpenAI's Decisions API in the same catalogue.
- Who is Microsoft-Decision-1 best for?
- Azure teams that need cheap, fast classification, routing or gating decisions under Microsoft's data terms, and will validate thresholds on their own labelled data.
- How much does Microsoft-Decision-1 cost for a team or in production?
- Input only, at $0.042 per million tokens: 10 billion tokens a month cost $420.
- Is Microsoft-Decision-1 safe for company data?
- Sold as a Foundry model by Azure, under which prompts are not used to train foundation models; DataZoneStandard keeps inference within a data zone, while GlobalStandard can process requests in any supported region.
- What are the best Microsoft-Decision-1 alternatives?
- The closest alternatives we have reviewed are Jev, Perplexity Decider (pplx-decider), GPT-6 Luna and Clef.
Answers checked Nov 6, 2026.