Behaviour for computing the cost of an LLM call.

Hardcoded prices go stale fast, so Baton does not ship authoritative prices. Instead it defines this behaviour and calls whichever module the host configures:

config :baton, pricing: MyApp.LLMPricing

Your implementation receives a usage map and returns a cost as Decimal.t(), or nil if the model is unknown (cost is then recorded as nil rather than a wrong number).

defmodule MyApp.LLMPricing do
  @behaviour Baton.Pricing

  @impl true
  def cost(%{model: "claude-sonnet-4-" <> _, input_tokens: i, output_tokens: o}) do
    Decimal.add(
      Decimal.mult(Decimal.new(i), Decimal.from_float(3.0 / 1_000_000)),
      Decimal.mult(Decimal.new(o), Decimal.from_float(15.0 / 1_000_000))
    )
  end

  def cost(_usage), do: nil
end

Baton.Pricing.Default is provided as a starting point and for tests, but you should supply your own and keep it current.

Service tiers

A step running in batch mode (use Baton.LLMStep, mode: :batch) spends roughly half as much per token, and its usage map says so with service_tier: "batch". Nothing downstream of here can recover that — by the time a cost is stored, a batch call and a live one look identical — so a pricing module that cares about the discount must branch on it:

def cost(%{service_tier: "batch"} = usage), do: discounted(usage)
def cost(usage), do: standard(usage)

Live calls carry no :service_tier key at all.