AI Rate Card

What a coding / tool-using agent costs a month on each model: tasks a day, model calls per task, context tokens per call, answer and reasoning tokens per call, percent of the context read from the prompt cache - priced at the cheapest real-time seller of each model (default: 13 usual coding models), cheapest first, USD, 30 days, with each model's reasoning class. Rows have card_url and cite.

What prompt caching and batch / flex save on one model: monthly input and output tokens priced at every listed seller plain, with the given share of input read from the cache (the seller's own cache-read price) and, if can_wait, the seller's batch / flex offers. USD, cheapest first; cache write fees are not counted. Each row has card_url and cite.

The cheapest offers for a whole monthly workload, ranked by monthly cost in USD: give a preset (a typical job; see list_workload_presets) and/or your own tokens a month. Best offer of each model; free promos and batch/flex/priority/regional prices left out unless asked. Each row: monthly_cost_usd, prices, context, checked_at, card_url and a ready cite line.

The n cheapest MODELS for a job (task): each model counted once at its own lowest current price (input tokens by default). One row per model with that price, where it is and card_url.

The cheapest current offers (rows: one model at one provider) of one model type, by input price. For the cheapest MODELS of a job (one row per model) use cheapest_models_by_task; for a monthly workload use cheapest_for_workload.

2 to 5 models side by side: the best current offer of each and, for up to 4, every field compared (lowest and median input/output price, context, max output, providers, regions, a 1M-in/200k-out workload cost), who wins by how many percent, providers selling all, and summary sentences.

The cost of a workload (input, output and cached input tokens, number of requests) at every current provider of one model, cheapest first, in USD - what the same job costs where.

One card with its citation. card_id is '<kind>:<key>': offer:<provider>|<model id as the provider writes it> (e.g. offer:anthropic|claude-opus-4), model:<creator>/<model> (e.g. model:anthropic/claude-opus-4), seller:<provider key> (e.g. seller:anthropic). Ids come from search_offers (offer_id, model_id, provider_id) and from the card_url of every result. Works for every seller, a hidden test seller too.

Every seller's current price of one model (an id, or a name to search for), cheapest input first.

AI Rate Card's articles on AI model prices (the cheapest models by task, price moves, costs), newest first, with their links - to quote or point a reader to.

The ready-made workloads of cheapest_for_workload: name, what job it is, tokens a month, context needed, model type.

The cost of an images / video / voice job at every model that sells it by its own unit, the cheapest offer of each model, cheapest first, USD: kind image (quantity = images), video (seconds), tts (characters to speak), stt (minutes of audio to transcribe). Models priced per token are not in it (use estimate_cost). Each row has card_url and cite.

The typical (median) current price of a job (task) across models, and the n models closest to it - what a normal price for this job is. Same arguments as cheapest_models_by_task.

Base models by class S/A/B/C/D in one direction (S best; today: reasoning), best score first: class (null = one source, score only), score 0-100, sources and each source's score (Arena, Epoch). A class needs two independent public sources. Method: https://airatecard.com/en/classes. Each row has card_url and cite.

The n most expensive MODELS for a job (task): each model counted once at its own lowest current price, highest first - the top of the market. Same arguments as cheapest_models_by_task.

What a RAG / document-search set-up costs in embeddings and reranking: pages of documents and tokens per page embedded once, the percent changed and re-embedded each month, queries a day embedded, and passages reranked per query - at the cheapest listed seller of every embedding and rerank model, USD, 30 days. The answer LLM is not included (use agent_cost / estimate_cost). Rows have card_url and cite.

Search current AI model offers (one model at one provider = one offer) by words, model type and limits; priced offers first, doubtful prices after sane ones. Each row has offer_id, model_id, provider_id, prices in USD per 1M tokens (or per unit), card_url and a cite line. Next: get_card {"card_id": "offer:<offer_id>"} for one offer, get_model_prices for one model everywhere, estimate_cost / cheapest_for_workload for a cost.

Is a chat subscription (e.g. Claude Pro) cheaper than the same use through the API? Messages a day and tokens per message priced at the cheapest real-time seller of a model and at the maker's own API, a month of 30 days, against the subscription's price, with the break-even messages a day. USD. preset claude-pro / mistral-pro takes the plan price from our plan tables; otherwise pass price.

For a provider (seller) joining the catalog: check a models.yaml before putting it online - the same check as the wizard and POST /v1/provider/validate. Give the file's text in `models` (YAML, or the same structure as JSON text); `domain` (the provider's site) makes the model_page checks exact. Returns ok, the problems (file, line, model, path, value, code, message, hint - fix each and call again), the catalog's preview of the models and next_action. No account needed for this check. Publishing needs a signed-in person: the provider_* tools (provider_start -> provider_verify_domain -> provider_check_files -> provider_publish) are on the account server /mcp/account (OAuth sign-in, scope provider), not on this public server; or the wizard at /join, or POST /v1/provider/applications. Templates: /static/templates/models.min.yaml (required fields only) and models.yaml (every field with hints).