microsoft decision-1 vs openai decisions api: which decision model should you use in 2026?

as of october 2026, the short answer to microsoft decision-1 vs openai decisions api is this: decision-1 has the lower list price per input token and, on microsoft's own tests, is the faster way to classify and route text. openai's decisions api is the one to pick when the input includes images, or when you need its documented data controls.

microsoft introduced decision-1 on october 9, 2026, three days after openai released its decisions api in beta. both bill only input tokens, and both answer with probabilities instead of prose.

our verdict: start with decision-1 for high volume text routing, and keep openai's endpoint for anything that has to look at a picture.

what did microsoft launch on october 9?

microsoft decision-1 is a decision model for routing, classification, prioritization, verification and workflow control. it does not write answers. you give it a fixed set of options, and microsoft says it returns a calibrated probability for each one. its stated bar: a 90% prediction should be right about nine times in 10 on representative cases. think of it as an ai routing model, not a chatbot.

microsoft says it built the model by post training qwen3.5-9b for fast, single pass scoring, and that it will soon rebase it on other models, including microsoft ai (mai) and openai models. it is available in microsoft foundry and through openrouter.

the foundry guide shows three question types:

• noul: a yes or no condition, returned as a probability from 0 through 1;
• choice: one option from your list, with a probability for every option;
• score: a value on an ordered scale, with a probability for each level.

the model returns numbers without a written rationale. a score on a four level scale, for example, runs from 0 through 3 and can land between levels. that is the point of an llm classification api like this: your code reads the answer and acts, with no parsing of free text.

how does microsoft decision-1 compare with the openai decisions api?

in our reading, the two products have nearly the same shape and differ on inputs, price and data controls. here is each vendor's own documentation side by side, checked on october 11, 2026.

openai says its endpoint returns typed answers about 10x faster than its responses api, and it expects general availability in the coming weeks. luna itself arrived in openai's api on september 22, and we compared it with anthropic's small model in our claude haiku 5.5 vs gpt-6 luna breakdown.

openrouter's listing for decision-1 shows a 32,768 token context and text as the only input type. that fits the foundry guide. in practice, a screenshot or product photo goes to openai's endpoint.

how much does microsoft decision-1 pricing save?

microsoft decision-1 pricing is lower per input token than openai's endpoint, and neither vendor bills output. with no output charge, the bill on both comes down to input tokens: what you send in each request.

openai's rate has fine print. its guide says regional processing premiums and long context input multipliers apply, and that the decisions rate covers /v1/decisions only. ordinary luna calls follow luna's normal model pricing, which our ai model pricing comparison, published before either decision api, lays out.

in our judgement, the price gap matters less than prompt discipline. a decision model that reads a whole ticket thread costs more than one that reads the last message, on either vendor. trim the input first, then compare rates.

are microsoft's benchmark claims believable?

microsoft's numbers look strong, and they are microsoft's. it says decision-1 had the highest accuracy in a 36 benchmark comparison spanning nearly 150,000 questions kept blind from training, and was the fastest model measured.

its headline speed claim names two models: 2.5x quicker than h2o lightning 4b v1.1, the runner up, and 35x quicker than gpt-6 sol. sol is a general model, so that 35x figure tells you more about choosing a decision model at all than about choosing between the two decision apis.

two other results are useful if you plan to automate on top of it:

1. robustness: microsoft perturbs each request in eight ways and says the decision flips on 1.3% of perturbations on average, with zero flips when options are shuffled or reversed.
2. safety: it tested 5,250 requests across 11 benchmarks covering harmful content, jailbreaks and prompt injection.

microsoft also lists openai's luna decisions among the models it benchmarked. until someone independent runs both on the same set, treat any ranking between them as a vendor claim. microsoft has also updated the post since launch to add benchmark results, so check the live page before you quote it.

what can a decision model actually do in an agent?

speed is the other reason to use one. microsoft's example: adding 100 milliseconds to each of 20 sequential decisions adds two seconds to a workflow.

in our view, a decision model is the cheap judge that sits between the expensive steps of an ai agent. microsoft's own suggestions include deciding whether an agent should continue, stop, retry or hand off, choosing which model handles a request, and grading an ai response to accept, revise or reject it.

microsoft gives an internal example: xbox research sorted more than 10,000 pieces of feedback into fixed themes and found decision-1 competitive with gpt-6 sol on quality, while running over 14x faster and at 200x less cost. microsoft also pitches it for picking the next interface action in computer use. in our reading, with text or json input, that means describing the screen as text first.

both vendors say where these tools stop. microsoft's guide says to use a generative model when you need to create, summarize or explain. openai points to structured outputs or function calling when you need your own schema or a tool call.

which decision model should you use?

our verdict: pick by input type first, then by price. as a rule of thumb:

• text routing, ticket triage, intent detection at volume: microsoft decision-1. lower rate, and microsoft's tests favor it on speed.
• anything with a photo or screenshot: openai decisions api, the documented option for image input.
• regulated data: openai, if you are an eligible customer for zero data retention or hipaa. on azure, check what your own agreement covers before sending sensitive records.
• you need a reason, not just a score: neither. use a generative model for the explanation, and a decision model only for the gate.
• decisions about people: never the sole decider. microsoft's guide asks for meaningful human review on credit, employment, housing, healthcare and legal matters.

our take: with microsoft and openai both selling decision endpoints, the router in front of your agent is now a cheap place to save money.

keep your context when the router changes

our guess: the model that routes your work this month will not be the one doing it next year. the part worth keeping is your own record of the work, not loyalty to one vendor.

remynd is a mac app that keeps that record. it captures the focused window, runs ocr locally with apple vision, and keeps the searchable index on your mac. you can exclude apps and sites, and history access is read only.

its agent window runs your own claude code or codex cli against that history.

download remynd for mac and keep your work history yours, whichever model makes the next call.