Usage and cost

Every model response records its token counts in the thread's log, so what a thread used and spent is read back from the log, not from a separate meter. Open the thread and ask it.

Tokens used

// A price on the model is what lets cost() count money (see Models).
const bot = agent({
  model: anthropic("claude-sonnet-5", {
    price: { input: 2000, output: 10000 },
  }),
});
const store = sqlite(".threads");

const result = await bot.run("Summarize the report.", { store });
const thread = result.thread;

const usage = await thread.usage();
if (!usage.ok) throw new Error(usage.error.message);
const { input_tokens, output_tokens, unknown_responses } = usage.value;

unknown_responses counts responses whose provider didn't report usage (for example a stream that broke), and the rare response whose counts would take a total past 2^53-1. They are never added as zero, so a non-zero count tells you the totals are a lower bound.

A fork counts its parent's prefix too: its usage is everything on its branch.

Money spent

cost() prices the recorded usage with the model's price. An agent whose model has a price records its currency as USD, so the model factory's price is in nano-dollars per token (USD 2 per million tokens is 2000, see Models).

const cost = await thread.cost();
if (!cost.ok) throw new Error(cost.error.message);
if (cost.value !== null) {
  const { currency, known_nanos, complete } = cost.value;
  console.log(`${currency} ${known_nanos / 1e9}`, complete ? "" : "(at least)");
}

cost() is null when none of the agent's models (its model and fallback) has a price: threads doesn't guess money it can't count. If only some have one, requests to an unpriced model make the cost incomplete (complete: false) rather than free.

A whole team

Pass tree: true to add every subagent's spend, at any depth, each read from its own log:

const total = await thread.cost({ tree: true });

The total never hides spend it can't count. It is in the root's currency, or the first priced subagent's; if no thread in the tree has a price it is null, which means unknown, not free. complete and bounded turn false when a subagent called a model with no price, or was priced in another currency, or when a subagent has no log to prove it spent nothing (for example one a cancel stopped before it started). If a subagent's log can't be read, the whole call returns that error, naming the subagent, rather than a partial sum.

Prompt-cache breaks

A cache break is a turn whose prompt-cache reads dropped sharply, which usually means you paid full price for a prompt that was cached a moment before. cacheBreaks() / cache_breaks() lists them with the likely cause:

const breaks = await thread.cacheBreaks();
if (!breaks.ok) throw new Error(breaks.error.message);
for (const b of breaks.value) console.log(b.request_event_id, b.likely_cause);

likely_cause is settings_changed (model or settings switched), compacted, context_edited, tools_changed, ttl_expired (the gap since the last request outlived the cache) or unknown. The cache lifetime it judges by is context.cache_ttl_ms. By default it is the lifetime your models declare, so anthropic("claude-sonnet-5", { promptCache: "1h" }) judges by an hour; see Cache lifetimes.

Errors

All three return a value instead of throwing. A log that fails its integrity check comes back as log_corrupt; one written by a newer threads as unsupported_format or unsupported_critical_event. cost() also returns cost_overflow when a total passes 2^53-1 nanos, rather than a rounded number.

Related