Reconcile a 9-clip batch: poll usage.cost versus /v1/usage debited_usd

Nine Seedance 2.5 clips at 720p, 9:16, 6 seconds cost $3.4668 each ($31.2012 if all nine complete). If two fail before capture, seven captures give $24.2676, and that is the number both the sum of poll usage.cost values and the ledger debited_usd should show.

When the two disagree, trust the ledger: the usage docs say never to sum rows yourself and to quote summary.debited_usd, which counts captured rows only.

What should reconcile

Reserved holds are open, not spent. Refunded rows are holds Sume gave back. Only captured rows are debits.

Compare both numbers in code

Poll each job, add usage.cost for completed ones, then ask the ledger for the same jobs.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def poll_cost(job_ids):
    total = 0.0
    for jid in job_ids:
        j = requests.get(f"https://api.sume.com/v1/videos/{jid}", headers=H, timeout=30).json()
        if j["status"] == "completed":
            total += j["usage"]["cost"]
    return round(total, 4)

def ledger_cost(job_id):
    r = requests.get("https://api.sume.com/v1/usage", headers=H,
                     params={"job_id": job_id, "limit": 50}, timeout=30)
    return r.json()["summary"]["debited_usd"]

Why they can differ

• A job still in_progress has a reserve but no capture, so the ledger shows it as held, not debited.
• usage.cost on a poll response is the Sume billable amount per job; summary.debited_usd also includes the agent's own turns if you scoped by thread or run.
• Query the ledger with job_id, run_id or thread_id; with a batch, tag jobs so you can scope one run.