Microsoft Decision-1 pricing

The published token price

The Microsoft price is for the model's API usage. The matching OpenRouter row is a provider listing checked separately on October 10. It does not establish that every account, region or purchasing agreement has identical terms. Check your selected service's billing settings before committing to volume.

Free output means the decision scores do not add an output-token charge at this rate. Input is still billable. A Microsoft 365 or Copilot subscription should not be treated as a bundle of Decision-1 API calls; the announcement supplies a usage price, not a subscription allowance.

Paying for a decision, rather than a written answer

Microsoft describes Decision-1 as a model post-trained from Qwen3.5-9B. You supply fixed answer options, and it assigns probabilities to them. Supported task shapes include yes/no checks, multiple-choice selection, ratings and grading against a rubric.

For a support queue, the options might be billing, account access and technical support. The result helps your application choose a queue. It does not write the customer response. If another model drafts that response, its usage belongs in a separate part of your budget.

This distinction matters when comparing prices. A small decision bill does not replace all the work in a chat application. Define the point where the score changes an action: accept a label, ask a reviewer, choose another model or stop a workflow. Include the cost of that action when assessing the design.

Request budgets you can reproduce

Use cost in USD = requests × average billable input tokens × 0.042 ÷ 1,000,000. If you already have total billable input usage, multiply that token count by 0.042 and divide by one million.

These are arithmetic scenarios, not measured invoices or savings claims. They assume the stated complete input on each call, no additional billed retries and the listed USD token rate. Taxes, currency conversion, platform charges and downstream services are outside the table.

Count the whole decision input

A ticket's text is only one part of a routing request. Instructions, answer descriptions and a grading rubric can make the actual input longer. Plan with the provider's billable usage record, rather than a character count for the ticket alone. The table's token figure is an assumption about the entire billed input.

For example, 100,000 requests averaging 4,000 input tokens cost $16.80. If billable retries add 2% to that same input volume, the total becomes $17.136, or about $17.14. The 2% is a planning assumption. Measure your own retry volume instead of assuming all failed requests are charged.

Foundry and OpenRouter access

Microsoft's current announcement links to Microsoft Foundry and OpenRouter's Decision-1 page. OpenRouter currently lists an Azure endpoint under microsoft/microsoft-decision-1, with text input and structured decision output. This verifies a public provider listing; we have not run a paid inference call.

OpenRouter identifies this as a Decisions API model and explicitly says ordinary chat-completions SDK calls will not work with it. Use that provider's decision request format. Do not paste the model name into an existing chat integration and assume that a zero output price makes it a free text generator.

Context and optional billing modes

OpenRouter lists a 32,768-token context window for its offering. Treat that as an OpenRouter limit, not a promise about every Foundry deployment. Request size and throughput limits are separate checks when planning a queue.

The estimates here use the listed input price throughout. They apply no cache, Batch, Flex or fast-mode discount. Do not carry a discount from a different model into this budget. If your chosen provider offers a separate mode or contract rate, calculate it from that provider's documented terms.

Budget for the rest of the workflow

Suppose 1 million 1,000-token decisions cost $42, and 5% of them call another service at an assumed $0.01 each. Those 50,000 downstream calls add $500, making the combined model-and-service estimate $542. Reducing escalations can matter more than a small difference in the first model's token price, but only if accepted decisions remain accurate enough.

Evaluate representative labeled examples before changing thresholds. Record accepted decisions, mistakes, escalations and latency alongside token usage. Microsoft's announcement reports internal benchmark results; those results do not establish your application's savings or error rate. A cheaper incorrect routing decision can create more work later.

For other structured-decision billing models, read our Perplexity Decider pricing guide and OpenAI Decisions pricing guide. Compare the same task and complete billed inputs. The site's general text-generation calculator does not model Decision-1's endpoint, so use the formula above for this estimate.

AI API Prices editors checked the linked first-party pages on October 10, 2026. Recheck the selected provider before purchase. Report a correction to [email protected].