Cost Tracking

Copy Markdown View Source

AgentSessionManager can estimate and persist per-run USD cost from token usage. The cost pipeline is model-aware and supports provider defaults, model prefixes, and cache-token pricing for Claude.

Overview

Cost tracking has four parts:

  1. AgentSessionManager.Cost.CostCalculator computes USD from token counts.
  2. Policy.Runtime uses model-aware rates for {:max_cost_usd, ...} enforcement.
  3. SessionManager stores run.cost_usd at successful finalization.
  4. Query and rendering surfaces expose cost in summaries/output.

Configuration

Set a pricing table globally:

config :agent_session_manager,
  pricing_table: %{
    "claude" => %{
      default: %{input: 0.000003, output: 0.000015},
      models: %{
        "claude-sonnet-4-5" => %{
          input: 0.000003,
          output: 0.000015,
          cache_read: 0.0000003,
          cache_creation: 0.00000375
        }
      }
    }
  }

If unset, ASM uses CostCalculator.default_pricing_table/0.

How Cost Is Calculated

CostCalculator.calculate/4 computes:

cost =
  input_tokens * input_rate +
  output_tokens * output_rate +
  cache_read_tokens * (cache_read_rate || input_rate) +
  cache_creation_tokens * (cache_creation_rate || input_rate)

Rate resolution order:

  1. Exact model match.
  2. Longest prefix model match (for dated model ids like ...-20250929).
  3. Provider default rate.

If no provider rates exist, calculation returns {:error, :no_rates}.

Policy Enforcement

Policy.Runtime captures model from :run_started events and applies model-specific rates for ongoing cost accumulation. If no model is captured, provider default rates are used.

Legacy flat policy_cost_rates remain supported for backward compatibility.

Querying Cost

Use QueryAPI.get_cost_summary/2:

query = {AgentSessionManager.Adapters.EctoQueryAPI, MyApp.Repo}

{:ok, summary} =
  AgentSessionManager.Ports.QueryAPI.get_cost_summary(query,
    provider: "claude",
    since: DateTime.add(DateTime.utc_now(), -86_400, :second)
  )

Returned keys:

  • total_cost_usd
  • run_count
  • by_provider
  • by_model
  • unmapped_runs

Runs without stored cost_usd are calculated post-hoc when possible.

Rendering and Telemetry

  • Compact/verbose renderers include cost labels when cost_usd is present.
  • JSONL compact sink includes a $ field on token_usage_updated.
  • Run-stop telemetry includes :cost_usd measurement when available.

Enterprise Pricing Tables

You can supply a custom pricing_table for negotiated rates, regional rates, or internal chargeback models. A per-query override is also supported in get_cost_summary/2 via :pricing_table.

Migration from Flat policy_cost_rates

Old format:

%{"claude" => %{input: 0.000003, output: 0.000015}}

New structured format:

%{
  "claude" => %{
    default: %{input: 0.000003, output: 0.000015},
    models: %{}
  }
}

CostCalculator.normalize_legacy_rates/1 accepts both formats during migration, so rollout can be incremental.