Platform

Pricing

What each version of A8 costs per million tokens, what counts as a token in a verdict, and how billing works.

This page as Markdown: https://u22a8.ai/docs/pricing.md. The index of every docs page: https://u22a8.ai/llms.txt.

A8 is metered per token. Each answered request is billed for the input tokens it reads, the reasoning tokens it writes, and the cached tokens a repeat reads from the cache, as the usage on its response reports them. The prices below are those of A8 1.1 Preview, the current version. Each version of A8 has its own prices.

§1Prices

ModelInput ($ / 1M tokens)Cached input ($ / 1M tokens)Output ($ / 1M tokens)
a8, your organization’s model$2$0$0
a8-1-1@reference, the reference model$2$0$0
Table 1. Prices of A8 1.1 Preview. Every update of your model, pinned or latest, and every minimum accuracy, is billed at these prices.
  • Input ($ / 1M tokens): the price of one million input tokens, summed over an organization’s requests in a billing period.
  • Cached input ($ / 1M tokens): a repeat of a request reuses its first reading, served from the cache, and usage reports that reading’s tokens as cached input. They are billed at this price, and the input price applies to the rest of the input.
  • Output ($ / 1M tokens): A8 scores the verdict, so its output tokens are reasoning tokens alone: what reading the request took with reasoning on. A request read as written, with reasoning off or already in A8’s layout, reports 0.
FeaturePrice
Teach$0
Retract$0
Fitting a new version$0
Precedents and recall$2 per 1M input tokens, the judging price
Table 2. Prices of the features beside judging.

§2What counts as a token

  • What A8 reads. A verdict counts the texts A8 judges with: the criterion, the subject and its standards, and each property’s own criterion. A recall counts the criterion it asks about.
  • A8’s own tokenizer. A8 counts tokens its own way, so the count for a given text differs from another provider’s count for the same text.
  • More than one choice. With n above 1 on Chat Completions, every choice holds the same verdict, and the input, cached and reasoning tokens each count once.
  • An abstention. An abstention is billed $0, its reasoning and cached tokens included. Its usage counts 0 input tokens besides them.

Each interface reports the counts in its own usage fields. On Chat Completions and Responses, the input count holds the cached tokens, so the input billed at the input price is the input count minus the cached count.

  • Chat Completions: prompt_tokens is the input count, prompt_tokens_details.cached_tokens the cached input, and completion_tokens the reasoning.
  • Responses: input_tokens is the input count, input_tokens_details.cached_tokens the cached input, and output_tokens the reasoning.
  • Anthropic Messages: input_tokens is the input billed at the input price, cache_read_input_tokens the cached input, and output_tokens the reasoning.
  • Decisions: input_tokens is the input count, with the input the questions share counted once, and output_tokens is 0, as the layout is read as written.

§3A worked example

A regression suite judges 10,000 replies against one criterion, with a verdict shape and reasoning off. Each reply runs about 300 tokens and the criterion with its shape about 100, so each request reads about 400 input tokens:

ItemAmount
Requests10,000
Input tokens per requestAbout 400
Input tokens in allAbout 4,000,000
Cost at the a8 priceAbout $8
Table 3. A worked example with illustrative token counts; your responses report the real ones. Requests A8 abstains on are billed $0.

§4Billing

Usage draws down prepaid credit. Prices are in US dollars, and accounts in Canada pay in Canadian dollars: CA$2 per 1M input tokens. An enterprise agreement sets its own price; write to [email protected].

Account holders are emailed at least 30 days before a price increase applies to them. The terms that govern billing are in the Terms of service.

Next steps