Pricing

Pay only when A8 gives a verdict

pay as you go

A8 1.1 Preview

per 1M input tokens

The tokens A8 writes while working out which part of a request is the criterion and which part is being judged. They are A8’s only output. With reasoning off, there are none. glossary: $0
When a request repeats, A8 reuses its first reading. The response reports that reading’s tokens as cached input tokens, in place of reasoning tokens. glossary: $0
A8’s reply when the evidence falls short of the minimum accuracy you set: it declines to give a verdict, and says so. glossary: $0
Sending A8 an example with the answer it should give. Your model learns it in its next update. glossary and Taking back a taught example. Your model drops it in its next update. glossary: $0

A model that learns from what you teach
The accuracy a verdict has to meet for A8 to give it: from 0 to 99%, set in the model name, as in a8:95, and 90% by default. Below it, A8 abstains. glossary from 0 to 99%, set per request
The taught examples behind a verdict, returned with it when you ask for them. glossary , at the same price per token

enterprise

Custom

an annual agreement

Volume pricing
Single sign-on and SCIM
Rate limits set for your traffic

options

An audit log of every verdict
Zero data retention for judged content

An abstention costs nothing

When the evidence falls short of the minimum accuracy you set, A8 abstains. The whole request is billed $0: its input, and any reasoning.

answered

"usage": {
  "prompt_tokens": 412,
  "completion_tokens": 0,
  "total_tokens": 412
}

abstained

"usage": {
  "prompt_tokens": 0,
  "completion_tokens": 0,
  "total_tokens": 0
}

Reasoning on or off, per request

Reasoning decides how A8 reads a request: which part is the criterion and which part is being judged. The judging is the same either way: the same model, the same minimum accuracy, the same abstentions.

how each API sets it

reasoning on · default

Reads any layout: an eval framework’s grader prompt, a criterion stated mid-conversation, a reference answer, or a follow-up about an earlier turn.

Billed as reasoning tokens at $0, and as cached tokens at $0 when the request repeats. A request already laid out for reasoning off uses none.

reasoning off

Reads the request as laid out: the system message is the criterion, and the latest user message is what is judged.

Uses no reasoning tokens, and answers faster.

Your data trains only your model

What you send trains your model and no other, and is never shared with another customer.

Withdraw a taught example in the console, and your model drops it in its next update.

how content is kept and deleted

questions

What counts as a token?: Input tokens are the text A8 reads: the criterion, what is judged, and any reference answer or context you send. Output tokens are reasoning tokens. A8 counts with its own tokenizer, so its count for a text differs from another provider’s. how tokens are counted
How do I see what I’ve used?: In the console, which shows your usage and remaining credit. Each response also reports its own count as usage.
Will these prices change?: Each version of A8 has its own prices. These are the prices of A8 1.1 Preview, the current version. Account holders are emailed at least 30 days before a price increase applies to them.
Does a higher minimum accuracy cost more?: No. Every minimum accuracy is billed at the same price, and so is every update of your model.

Prices are in US dollars. The Terms of service govern billing.