Decision models explained: Microsoft-Decision-1, OpenAI's Decisions API and when to use them

Two of the biggest AI platforms launched the same new kind of product within a week. On October 6, 2026, OpenAI released its Decisions API in public beta. On October 9, Microsoft released Microsoft-Decision-1, a small model post-trained from Alibaba's Qwen3.5-9B. Neither writes prose. You give them an input and a fixed set of possible answers, and they return a probability for each answer that your code can act on. This explainer covers how decision models work, how the two offerings differ, and where they fit.

It is based on OpenAI's and Microsoft's documentation, Microsoft's announcement and OpenRouter's model listing, all checked October 12, 2026.

The short version

• What a decision model does: scores answers you define ("billing", "engineering", "support"; yes/no; a 1–4 severity scale) and returns a probability for each, in one pass, with no generated text.
• Microsoft-Decision-1: a hosted model (no open weights) on Microsoft Foundry and OpenRouter. It costs $0.042 per million input tokens, and output is free.
• OpenAI Decisions API: POST /v1/decisions with gpt-6-luna, in public beta. It costs $0.10 per million input tokens, input only. OpenAI says it is "10x faster than the Responses API."
• Good for: classification, routing, verification, guardrails and grading other models' output.
• Not for: chat, summarization, translation or anything that needs an explanation.

What is a decision model?

A chat model answers "Which team should handle this ticket?" by generating text, perhaps a sentence, perhaps the word "engineering". Your code then has to parse that text, and you get no sense of how sure the model was.

A decision model inverts that. You send the input and a list of allowed answers with short descriptions. The model reads everything once and assigns a probability to each answer. Microsoft's Foundry documentation gives this example: the input "The API returns 500 on every call", with options for billing, engineering and support. The result is the chosen option plus a probability for every option, as a typed object your code can branch on.

Microsoft's announcement describes decision models as purpose-built for "structured outputs that software can act on" rather than text generation. Because nothing is generated token by token, a call is fast and cheap. That is why output is free or not billed in both offerings.

Three question shapes

Both products support the same three basic shapes, under slightly different names:

OpenAI calls the yes/no type predicate. Both also use choice and score. You can ask several independent questions about the same input in one request. OpenAI's guide says decisions that depend on each other need separate requests.

What "calibrated" means

Microsoft says Decision-1 returns a calibrated probability for each option. Calibration means the numbers match reality: across many cases where the model says 90%, it should be right about nine times in ten. Microsoft states that as its goal. The announcement doesn't publish a full calibration curve, though MarkTechPost reports a calibration score Microsoft gave elsewhere (see the results section below).

Calibration is what makes a probability useful. You can set a threshold, for example "auto-route above 0.85 and send everything else to a person," and know roughly what error rate you're accepting. OpenAI's guide makes the same point from the other direction. It tells developers to "use labeled examples from your application to set thresholds," weighing the cost of false positives against false negatives. Neither vendor promises its probabilities are calibrated for your data.

Microsoft-Decision-1 vs OpenAI's Decisions API

Last verified: October 12, 2026. Neither API is compatible with chat-completions SDKs. OpenRouter notes that Decision-1 runs on its separate Decisions endpoint.

Two practical differences stand out. OpenAI accepts images, so you can ask a yes/no question about a product photo. Decision-1 is text-only. And Decision-1's weights change over time behind a stable API, which is convenient but means a threshold you tuned in October may need re-checking later.

What are decision models good for?

Classification and labeling

This is the obvious use: sorting tickets, tagging feedback, filtering content. Microsoft says Xbox Research used Decision-1 to label more than 10,000 pieces of game feedback with quality "competitive with" GPT-6 Sol, more than 14 times faster and 200 times cheaper. That is a vendor-reported internal case study, not an independent test.

Routing

Routing is classification where the labels are destinations: which team, which tool, which workflow or which model. Our explainer on model routing describes the "classifier router" pattern, a cheap model that decides up front whether a request needs a small, mid-tier or frontier model. A decision model is a natural fit for that first step. It returns a probability for each tier, so you can also escalate whenever the router itself is unsure.

Verification and guardrails

This may be the most important use for agents. Before an agent takes an action, a decision model can answer narrow questions: Is this action within the user's request? Is this answer grounded in the supplied evidence? Does this output break a policy? Microsoft's Foundry model card lists grading AI responses and proposed agent actions, and groundedness checks, among Decision-1's tasks.

That matches the advice Microsoft's own CEO gave on October 10: don't let a single model check its own work (see Satya Nadella's 'emergency brake' principles). An independent, cheap scorer is one way to add that check without doubling your bill. It is not a sandbox, though. Our case against unattended agents explains why hard limits on what an agent can touch still matter.

Grading and evaluation

The score type turns rubric-based grading ("rate this answer 0–3 for completeness") into a single fast call. Microsoft says its Copilot team uses Decision-1 for quality control and found it competitive with GPT-5.6 Luna and 100 times faster, a vendor-reported claim.

Third-party and vendor-reported results

So far, the numbers come from the vendors:

• Microsoft (vendor-reported, announcement dated October 9, 2026): Microsoft says Decision-1 had the highest accuracy in its 36-benchmark comparison with nearly 150,000 questions, which it says were kept out of training. MarkTechPost reports the figures as 83.5% average accuracy, against 81.9% for the next model, Quyet-1.0-Large. Microsoft reports median latency about 35 times faster than GPT-6 Sol, and a 1.3% average "decision-flip" rate when it perturbed inputs eight different ways, with zero flips from shuffled or reworded options. Microsoft edited the post on October 10 to add results from JevBench, a third-party decision-model leaderboard. Accounts of the runner-up's speed differ between versions of the post (MarkTechPost cites 4.5x faster than Quyet-1.0-Large; the current post cites about 2.5x faster than H2O-Lightning-4B). MarkTechPost notes that H2O.ai disputes JevBench's latency method and that independent results are pending.
• OpenAI (vendor-reported): "10x faster than the Responses API." The guide gives no benchmark setup for that figure.

Treat all of these as claims to check on your own data. The accuracy numbers average many benchmarks that may not resemble your task, and the latency comparisons pit a 9-billion-parameter scorer against general-purpose models doing a different job.

Limitations

Microsoft's documentation is candid about the weak spots:

• Wording matters. Results change with how you phrase the question and options, and even with option order. Microsoft suggests randomizing option order.
• No explanations. You get numbers, not reasons, which makes debugging and auditing harder.
• Familiar tasks only. Calibration is strongest on task types like those the model was trained on.
• Force-choice risk. Without an explicit "cannot tell" option, the model must pick something. Microsoft recommends adding one, and escalating when the top probability is low.
• Not for consequential decisions on their own. Both Microsoft's model card and documentation say not to use scores as the sole basis for decisions about credit, employment, housing, healthcare or legal rights.

How to start

1. Pick one narrow, high-volume decision you already make with a chat model or by hand.
2. Write the options with short, distinct descriptions, and include a "cannot tell" or "other" option.
3. Run a few hundred labeled examples through it and choose a threshold from the results, as both vendors recommend.
4. Route low-confidence cases to a person or a stronger model, and re-check the threshold when the vendor updates the model.

What's confirmed, reported and uncertain

• Confirmed (vendor documentation): release dates; endpoints; question types; prices; Decision-1's base model, text-only input and limitations; OpenAI's beta status, image support and gpt-6-luna-only model support.
• Vendor-reported: all accuracy, calibration, latency and cost-saving figures, including the internal Xbox, Copilot and Discovery case studies.
• Uncertain: independent accuracy and calibration results; when OpenAI's API reaches GA; what changes when Microsoft rebases Decision-1 on MAI or OpenAI models.