OpenAI Decisions API vs Microsoft Decision-1 | Pricing and Cost per Decision

Most production AI calls are not essays. They are small judgments inside software: which queue does this ticket belong in, is this message spam, does this answer follow the policy, should the agent be allowed to run this command. Teams have been paying a general-purpose language model to write out an answer like "billing" and then parsing it. Decision models remove the writing step. You send the evidence and a fixed set of options, and the model returns numbers your code can act on directly.

Two of these launched within days of each other. OpenAI released the Decisions API in public beta on 6 October 2026, and Microsoft introduced Microsoft-Decision-1 on 9 October 2026. This guide compares them on the thing that decides most real projects: what a decision actually costs at volume. Keep the pricing database open if you want to check the general-purpose models we compare against.

What a decision model is, in plain terms

A normal chat model generates tokens one at a time until it decides to stop. A decision model scores a small set of answers you supply and returns them in a fixed shape. OpenAI describes three question types: a predicate that returns the probability a condition is true, a choice that picks one of your options, and a score that rates the input against ordered levels. Microsoft's documentation describes the same idea with yes or no, multiple choice, rating and rubric questions.

Two practical consequences follow. First, there is no output text to pay for, which is why both vendors price on input only. Second, you get a probability instead of a word. That lets you set a threshold, for example auto-route above 0.85 and send everything else to a person, which is hard to do reliably with free-form text.

The rate cards side by side

Per token, Decision-1 is about 58% cheaper. Microsoft has not published a separate Foundry per-token rate in its announcement, so the OpenRouter figure is the one we can verify. If you buy through Azure under an enterprise agreement, check the price on your own Foundry deployment before budgeting.

What one decision actually costs

Because output is free, the cost of a decision is simply the number of input tokens multiplied by the input rate. The input includes your evidence, your instructions and the list of options. A realistic support ticket with a short instruction block and five queue names is around 600 tokens. You can measure your own with the token calculator.

At small volumes the difference is lunch money, and the choice should come down to accuracy and integration effort. The gap starts to matter at tens of millions of decisions, or when each decision carries a lot of context.

Worked example: checking agent actions

Suppose a coding agent proposes about 50 tool calls per task, your team runs 4,000 tasks a month, and before each call you ask a decision model whether the action stays within policy. Each check includes the policy text, the recent conversation and the proposed command, roughly 1,500 tokens. That is 200,000 checks and 300 million input tokens a month.

• Decisions API: 300 × $0.10 = $30.00 per month.
• Decision-1: 300 × $0.042 = $12.60 per month, but only if every check fits in 32,768 tokens.
• For comparison, the agent's main work on a frontier model will usually cost far more than either, so the guard rail is a small line item.

The lesson is that decision models are cheap enough to add to every step. The question is less about price and more about whether the probabilities are trustworthy on your data.

When a normal model with caching is cheaper

This is the point most launch coverage skips. The Decisions API charges $0.10 for every input token, including the instruction block you send with every request. The same gpt-6-luna model on the Responses API costs $0.10 input, $0.01 for cached input and $0.50 output per million tokens. If your instructions are long and identical across calls, caching can outweigh the free output.

Two caveats keep this honest. The Responses figure assumes the model returns a short label of about 20 tokens; if reasoning or a longer answer is produced, output cost rises quickly. It also assumes the cache hits consistently, which depends on prompt structure and traffic patterns. What you lose on that route is the calibrated probability and the speed. OpenAI says the Decisions API returns answers about 10 times faster than the Responses API, a vendor figure worth checking on your own requests.

Speed and accuracy claims, and how to read them

Microsoft says Decision-1 had the highest accuracy across its 36-benchmark comparison of nearly 150,000 questions and was 35 times faster than GPT-6 Sol in its testing. OpenAI says the Decisions API is about 10 times faster than its Responses API. None of these figures had been independently reproduced when we checked, and each vendor chose its own comparison set.

Early developer reports are mixed, which is normal for a new category. Some testers in OpenAI's community thread said other decision models handled nuanced judgment calls better on their data. One independent write-up on the Go SDK also found that a refused answer can decode as a probability of zero, which a threshold check would treat as a confident no. Handle refusals explicitly before you put either API in front of a routing rule.

Pros and cons

OpenAI Decisions API

• Pro: accepts images, useful for damage checks, document triage and screenshot review.
• Pro: same vendor, billing and data controls if you already use OpenAI, with Zero Data Retention and regional processing for eligible customers.
• Pro: no output charges, so cost is predictable from input size alone.
• Con: no cache discount, so long repeated instructions are billed in full every time.
• Con: public beta on a single model, so behaviour and limits may change before general availability.

Microsoft-Decision-1

• Pro: lowest verified per-token price of the two, about 58% below the Decisions API.
• Pro: available through OpenRouter as well as Foundry, so you are not tied to Azure.
• Pro: built on an open-weight base model, which Microsoft says it will later rebase on other models.
• Con: text and JSON only, with no image input.
• Con: 32,768 token context, which rules out long documents or long agent histories in a single check.

Which one to choose

• Classifying text at high volume with short inputs: start with Decision-1 for price, then confirm accuracy against the Decisions API on the same sample.
• Anything involving images: the Decisions API is the only one of the two that accepts them.
• Long, identical instruction blocks and no need for probabilities: price out gpt-6-luna on the Responses API with caching first.
• Regulated data already governed under an OpenAI or Azure agreement: pick the one your existing contract covers, since the price gap is small at moderate volume.

If you are still deciding which general-purpose model sits behind the decision layer, the Cheapest LLM API Finder ranks current models by estimated monthly cost for your workload, and the comparison tool puts any two side by side.

How to test before you switch

1. 01Collect 300 to 500 real items your current system has already handled, with the correct answer recorded by a person.
2. 02Write the question and options once, and send the identical set to each option you are considering.
3. 03Record accuracy, the share of answers above your intended confidence threshold, latency and input tokens.
4. 04Check how each API reports a refusal or an unanswerable item, and make sure your code handles it.
5. 05Calculate cost per correct decision, not cost per token, and repeat the test once the Decisions API leaves beta.