Tokens & cost

agtail surfaces token usage and an approximate USD cost at two granularities:

• Per turn — each assistant turn (one API call) gets a 42,548 tok ≈$0.5851 badge, in both the CLI show and the web timeline.
• Per session — stats and the web timeline header show the session totals (input / output / cache-read) and the summed cost.

Tokens are consumed by the assistant turn, not by individual tool calls — a tool call's cost is attributed to the turn that emitted it. Claude reports usage on each assistant message; Codex's token_count event is attached to the preceding response.

Prices come from LiteLLM

Rather than hand-maintaining a price table, agtail sources model prices from LiteLLM's community-maintained sheet, model_prices_and_context_window.json.

• The file is fetched once and cached on disk at ~/.cache/agtail/litellm_prices.json.
• It is refetched when the cache is older than 7 days.
• Offline with a stale cache, the cached prices are used; with no cache at all, cost shows as unknown and a note is printed to stderr.

Model ids are matched by exact id, by the part after a provider /, and with a trailing -YYYYMMDD date stripped (treated as the same model).

Unknown cost is deliberate

If LiteLLM doesn't list a model, agtail shows "cost unknown" instead of guessing a price — tokens are still shown exactly. This follows the project's no-fallback principle (Concepts): a wrong number that looks authoritative is worse than an honest "unknown".

If a session mixes a priced and an unpriced model, the session total is reported as unknown (it would be incomplete), and the UI names which model wasn't priced.

Accuracy notes

• Costs are approximate. Cache-read and cache-write tokens are billed differently from fresh input; agtail applies LiteLLM's per-token rates for each where available.
• Each assistant turn's input includes the full (largely cached) context, so summing per-turn input across a long session is expected to be large — that is what you are billed, with cache discounts applied.