Milliseconds.ai

Milliseconds.ai covers a small set of decision types rather than one general one. You can get a yes/no answer, a classification label with a full score distribution, a rating on a scale, an answer drawn from the source text, extracted fields, named entities, or a verification result. Each shape fits a job that software already does somewhere, and the choice between them usually comes down to whether your code needs a flag, a category, a number, or a set of values it can store. A routing rule needs a label. A form-filling step needs fields. Picking the right shape is most of the setup work.

The model reads the input in one forward pass and returns the result, instead of writing an answer word by word like a chat model. That's why a decision comes back in roughly 90 milliseconds. Longer calls still exist. Answer runs around 0.3 to 0.9 seconds and extract takes 2.5 to 3.5 seconds. For a request that just needs a flag or a label, the speed holds up.

When a classification is close, you see it. A classify call returns the probability for every label you supplied, not just the winner, along with a confidence value. If two labels land near 0.34 each, your code can flag the case for review instead of guessing. That's a real difference from APIs that hand you a single answer and hide the doubt. No guessing required.

The same capabilities take text or images. Send bytes, a base64 string, or a data URL, and the model reads it directly. Supported formats are JPEG, PNG, and WebP, capped at 5 MB each. What about image storage? There's none. Image handling is billed as a fixed token count based on the detail level you pick. Remote image URLs aren't accepted, so your code has to send the file itself.

There are official TypeScript and Python SDKs, plus a command-line tool called dm1. The result types follow the labels you define, so you can read the winning label and its score without unpacking raw HTTP. The CLI reads text from a file or a pipe, writes JSON for another program, and uses exit codes to gate a script. Install is a single npm or pip command.

Production pricing is $0.04 per million input tokens, and output tokens are free. A test key gives you 125 million free input tokens every month, which is enough to evaluate the API on real data before paying anything, and a credit purchase of at least $10 is all it takes to switch a project over to production keys. Images bill at a fixed rate per detail level, so you can estimate cost per thousand images before you commit.