Cost & tokens

Breadcrumb tracks token usage on every LLM span and computes cost from your price table. It ships no prices of its own — if you don’t declare a model’s price and don’t set a cost yourself, cost stays null rather than being guessed.

Token types

Usage is normalized to the OpenTelemetry convention: inputTokens is always cache-inclusive, and the finer tiers are subsets of it.

Field	Meaning
`inputTokens`	Total prompt tokens, including cached reads and writes.
`outputTokens`	Total completion tokens, including reasoning.
`cachedInputTokens`	Cache-read tokens (a subset of input), billed at a discount.
`cacheWriteTokens`	Cache-creation tokens (a subset of input), billed at a premium.
`reasoningTokens`	Thinking tokens (a subset of output).

Breadcrumb reconciles the conflicting conventions of the AI SDK (ai.*), OpenTelemetry (gen_ai.*), and manual spans, so these come out consistent no matter the source.

Declaring prices

Prices are USD per 1M tokens. cachedInput and cacheWrite default to the base input rate when omitted — so no discount or premium is ever invented.

breadcrumb({
  pricing: {
    "gpt-5": { input: 1.25, output: 10, cachedInput: 0.125 },
    "claude-sonnet": { input: 3, output: 15, cacheWrite: 3.75 },
  },
});

Keys match as lowercase substrings of the span’s model, longest key wins — so "claude-opus" beats "claude". Cost is split across non-cached input, cache-read, cache-write, and output tiers, then summed.

The Cost view

The built-in dashboard’s Cost tab charts spend per day and model, and breaks it down by function — the functionId the call was given, falling back to the run’s root-span name for spans that carry none — so you can see which flow is expensive. Everything there is also available programmatically:

const summary = await bc.api.costSummary({ days: 14, environment: "production" });