Advisor (opus) token costs not tracked — data is in iterations[].type=advisor_message #1115

Summary

When Claude Code's advisor() tool is used (e.g. sonnet as primary model, opus as advisor), the advisor model's token costs are not counted by ccusage. The data is present in the JSONL — it's embedded in the message.usage.iterations array as entries with type: \"advisor_message\" — but ccusage doesn't parse the iterations array, so those tokens are silently dropped.

Where the data lives

Each type: \"assistant\" entry in the session JSONL has a message.usage.iterations array. When the advisor tool fires, a middle iteration of type: \"advisor_message\" is inserted between the two main-model iterations:

{
  "type": "assistant",
  "advisorModel": "claude-opus-4-7",
  "message": {
    "model": "claude-sonnet-4-6",
    "usage": {
      "input_tokens": 2,
      "output_tokens": 491,
      "cache_read_input_tokens": 226584,
      "cache_creation_input_tokens": 7853,
      "iterations": [
        {
          "type": "message",
          "input_tokens": 1,
          "output_tokens": 45,
          "cache_read_input_tokens": 109696,
          "cache_creation_input_tokens": 7192
        },
        {
          "type": "advisor_message",
          "model": "claude-opus-4-7",
          "input_tokens": 159419,
          "output_tokens": 7805,
          "cache_read_input_tokens": 0,
          "cache_creation_input_tokens": 0
        },
        {
          "type": "message",
          "input_tokens": 1,
          "output_tokens": 446,
          "cache_read_input_tokens": 116888,
          "cache_creation_input_tokens": 661
        }
      ]
    }
  }
}

Key observations from the above (sampled from a real session, 2026-05-22T02:34:40):

• The advisor_message iteration carries its own model field (claude-opus-4-7) — the pricing is distinct from the main model
• cache_read_input_tokens: 0 — the full transcript is forwarded uncached to the advisor, so these are billed at full input rate (not the cheaper cache-read rate)
• The top-level cache_read_input_tokens: 226,584 is the sum of both main-model iterations only (109,696 + 116,888) — the advisor's 159K input tokens are separate and currently invisible to any external cost tool

Cost impact

In the example above:

• 159,419 uncached input tokens at opus pricing disappear entirely from ccusage's accounting
• 7,805 output tokens from opus also untracked

Since the advisor forwards the full conversation transcript each time it's called, these counts grow with session length. On a long session this can represent a substantial fraction of actual billing.

Proposed fix

When processing an assistant entry, check message.usage.iterations for any entries with type === \"advisor_message\". For each such entry:

1. Read model from the iteration (e.g. claude-opus-4-7)
2. Attribute input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens to that model using its own pricing
3. Do not also count those tokens toward the main session model — they are already excluded from the top-level usage totals

The advisorModel field at the top level of the JSONL entry is a reliable sentinel that the iterations array will contain an advisor_message entry, if you want to skip parsing iterations on unaffected turns.

Related

• advisor() tool inflates reported input tokens by forwarding full transcript, triggering premature auto-compaction on extended context models anthropics/claude-code#53065 — describes how the top-level cache_read_input_tokens sums both main-model iterations, making context appear ~2× its actual size (separate but related data-quality issue in the same field)
• [BUG]Advisor tool injects same-tier opus calls into every opus subagent, doubling spend invisibly anthropics/claude-code#58217 — advisor spend invisibility from the Claude Code side