claude -p total_cost_usd vs your Sume bill: which number to trust

total_cost_usd in the claude -p --output-format json result is Claude Code's own figure for the model usage in that session. It is not your Sume bill. Generation that your agent starts through the Sume hosted MCP server is billed by Sume, and GET /v1/usage is the record for it, so a script that tracks spend needs to read both.

Two meters, two owners

Claude Code's headless docs say that with --output-format json the payload includes total_cost_usd and a per-model breakdown, so a scripted caller can track spend without opening a dashboard. That is useful for the model side of an unattended run. It cannot see what the Sume tools charged, because those charges are made on Sume's servers when a generation job is admitted.

What Sume counts

Sume's own receipts follow the same split. On Format runs and scheduled runs, usage.billable_amount_usd_micros covers generation spend only and leaves out the agent's LLM turn. The docs call GET /v1/usage the authoritative billing record, and a figure on a receipt is a receipt value, not an invoice.

Log both numbers after a headless run

The shell script below runs one tiny headless call against the hosted server, then prints both numbers. It builds the MCP config with jq so your key never lands in a file you commit, and it assumes claude, jq and curl are installed and SUME_API_KEY is set. The prompt only calls mcp_health, a read tool, so it generates nothing and spends nothing on Sume.

jq -n --arg k "$SUME_API_KEY" '{mcpServers:{sume:{type:"http",url:"https://mcp.sume.com/mcp",headers:{Authorization:("Bearer "+$k)}}}}' > /tmp/sume-mcp.json
claude -p "Call the mcp_health tool and print the result." \
  --mcp-config /tmp/sume-mcp.json --output-format json > /tmp/run.json
echo "claude side:"
jq -r '.total_cost_usd' /tmp/run.json
echo "sume side:"
curl -sS https://api.sume.com/v1/usage \
  -H "Authorization: Bearer $SUME_API_KEY" | head -c 600

Where to put the limit

Run the script once with a read-only prompt to see the first number, then once after a capped paid call to see the second move. Keep the two separate in your logs. If you add them, label the sum as an estimate.

For paid calls, do not rely on either meter to stop a loop. Put the limit on the request: max_spend_usd on an MCP paid tool, generation_spend_cap_usd on a Format run, and an idempotency_key on every create so a retry does not bill twice.

Short version

The two numbers answer different questions, and each is accurate for its own question. Read the first for the Claude side, read the second for what Sume charged, and trust the ledger when they disagree with a receipt.