> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.milliseconds.ai/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.milliseconds.ai/_mcp/server. # decision-machine-1 `decision-machine-1` is a decisions API. You send text and a question about it. You get a typed answer back: a yes/no, a label, a path through a label tree, a rating, an answer span, a filled JSON Schema, a list of entities, or a value check. It does not write prose. It does not chat. Every call returns structured JSON with probabilities, so your code can branch on a number instead of parsing a sentence. ## Watch the walkthrough Nina, who leads applications development, walks through what `decision-machine-1` does in under six minutes. She covers invoices, leases, clinical notes, support tickets and a pre-check in front of a chat model. Every call goes to `https://api.milliseconds.ai` with your milliseconds.ai API key. See [Authentication](/authentication). The price is \$0.04 per million input tokens and \$0 per output token. See [Pricing](/reference/pricing). > **Note** > > **Decisions, not generation.** Each capability answers one question about one text. Most calls finish in about 0.4 seconds. ## The whole surface | Endpoint | Question it answers | Returns | Typical latency | | ------------------------------------------- | --------------------------------------------------------------------------- | -------------------------------------------------------------------- | ----------------- | | `POST /v1/decision-machine-1/yes-no` | Is this statement true of the text? | `statement`, `answer`, `probability` | 0.42 s | | `POST /v1/decision-machine-1/classify` | Which label applies? | `label`, `probability`, `confidence`, `scores` | 0.43 s | | `POST /v1/decision-machine-1/classify-tree` | Which path through this label tree applies? | `path`, `label`, compound `probability` and `confidence`, `levels[]` | 0.77 s (2 levels) | | `POST /v1/decision-machine-1/rate` | Where does this sit on a scale? | `score`, `level`, `confidence`, `scores` | 0.42 s | | `POST /v1/decision-machine-1/answer` | What does the text say about this? | `question`, `answer` span, `probability`, `start`, `end` | 0.40 s | | `POST /v1/decision-machine-1/extract` | Fill this JSON Schema from the text. | `data` | 0.41 s | | `POST /v1/decision-machine-1/entities` | Where is every span of each type? | `entities` with type, text, probability, offsets | 0.42 s | | `POST /v1/decision-machine-1/verify` | Does the text support this value? | `matches`, `probability`, `found` | 0.40 s | | `POST /v1/chat/completions` | OpenAI clients: extraction and function calling | Chat completion object | 0.39 – 0.72 s | | `POST /v1/systemone` | TypeSafe clients: named `noul`, `choice` and `score` questions on one state | One typed answer per question | 0.17 s | | `GET /v1/models` | Which models does the OpenAI-compatible surface serve? | A list with one model object | — | Every capability accepts `text` (up to 20,000 characters) or `texts` (up to 32 items). A batch returns `{ "results": [ ... ] }` in input order. The latencies are wall-clock times against production on 2026-09-18, median of 7 runs from a laptop. They include TLS and internet transit. ## One call ```bash curl -X POST https://api.milliseconds.ai/v1/decision-machine-1/classify \ -H "Content-Type: application/json" \ -H "authorization: Bearer test_sk-..." \ -d '{ "text": "I was charged twice for my subscription this month and support has not replied.", "labels": { "billing": "payments, invoices, charges, refunds", "shipping": "delivery, tracking, returns", "account": "login, password, profile settings" } }' ``` ```json {"label":"billing","probability":0.995,"confidence":0.972,"scores":{"billing":0.995,"shipping":0,"account":0.005}} ``` The `label` picks the branch. The `confidence` decides whether a person reviews it. ## An OpenAI-compatible surface Point any OpenAI client at `https://api.milliseconds.ai/v1` with model `decision-machine-1`. Two modes work: `response_format` with a `json_schema`, and `tools` for function calling. Set the client's `api_key` to your test or production key. Plain chat returns `400 unsupported_request` on purpose. The model decides, it does not chat. TypeSafe clients keep their request shape too. `POST /v1/systemone` takes one `state` and named `noul`, `choice` and `score` questions. See [TypeSafe-compatible API](/typesafe/overview). #### [All capabilities](/capabilities/overview) The eight endpoints, their fields and their routing rules. #### [OpenAI-compatible API](/openai/overview) Reuse an OpenAI SDK for structured extraction and function calling. #### [TypeSafe-compatible API](/typesafe/overview) Send systemone requests with noul, choice and score questions. ## When not to use it * **You need generated text.** Summaries, rewrites, replies and explanations need a large language model. * **You need multi-step reasoning.** The model makes one pass over the text. It does not plan or chain steps. * **You need a conversation.** There is no chat mode and no memory between calls. * **Your task has no shape.** Each capability needs your labels, statements, questions or schema. You define the decision. * **You need open-ended discovery.** `entities` finds the types you name. It does not invent new ones. For the rest, a typed decision beats a prompt: no parsing, stable probabilities, and cost linear in characters. ## Next * [Quickstart](/quickstart) — one curl call to your first decision. * [Why a decisions model](/why-a-decisions-model) — where this beats an LLM prompt, and where it does not. * [Choosing a capability](/concepts/choosing-a-capability) — a decision table for the eight endpoints.