All capabilities

Seven typed decisions over text, the question each one answers, and how fast it returns.

Every capability is one POST to https://api.milliseconds.ai, with text or texts in the body. No API key is required during the launch period.

The seven capabilities

CapabilityQuestion it answersBacking modelTypical latencyEndpoint
Yes / noIs this statement true of the text?Classifier0.75 s – 1.25 sPOST /v1/decision-machine-1/yes-no
ClassifyWhich one label applies?Classifier1.18 s (3 labels)POST /v1/decision-machine-1/classify
RateWhere on this scale does the text sit?Classifier1.02 s (4 levels)POST /v1/decision-machine-1/rate
AnswerWhich span of the text answers this question?Extractor0.56 sPOST /v1/decision-machine-1/answer
ExtractWhat fills this JSON Schema?Extractor0.72 s (9 properties)POST /v1/decision-machine-1/extract
EntitiesWhere is every span of these types?Extractor0.50 sPOST /v1/decision-machine-1/entities
VerifyDoes the text support this value?Extractor0.48 sPOST /v1/decision-machine-1/verify

Latencies are wall-clock curl measurements from a laptop on 2026-09-16, over short inputs of 55 to 198 characters. They include TLS and internet transit. The two model families set the split: classifier calls land near 1 s, extractor calls near 0.5 s. How it works names the two models and gives the measured capacity table.

The routing rules

  1. You need a true or false verdict on one claim: use yes-no.
  2. You need exactly one label from a fixed set: use classify.
  3. You need an ordered level, such as severity or sentiment: use rate.
  4. You need a value that the text states, with its offsets: use answer for one question. Use extract for a whole record, or entities for every occurrence.
  5. You already hold a value and must check it against the text: use verify.

Choosing a capability works through the near-miss pairs, such as yes-no against classify with two labels.

What every capability shares

Each capability response carries two headers: x-input-chars, the number of input characters, and x-input-tokens, the input tokens billed for this call. Cost scales with everything you send: the text plus the labels, statements, questions, or schema. See Pricing.

Writing the labels and statements

The label text is what the model reads. "yes" and "no" carry no meaning for it. Write text that describes the case.

A when_true of "yes" with a when_false of "no" measurably flipped a correct yes-no answer to wrong. The probability fell from 1 to 0.004 on the same input. Writing good statements and labels shows the wording that works.

Try each one

Open the API Reference tab and use the Try it panel on any endpoint page. The panel needs no key and no setup. The Playground guide gives a worked example per capability and names what to read in each response.

Next