Check your Sume balance before a 100-clip batch: Python preflight

Before you queue a batch, read the workspace balance and compare it with the batch estimate. Sume's balance endpoint returns available_amount_usd_micros, the spendable amount in USD micros, so the comparison is integer arithmetic.

The estimate

The script assumes 100 clips of 5 seconds each at Wan 3.0 720p, where one clip is $0.625 billed, so the batch needs $62.50 of holds if every job is in flight at once. Change the two constants for another model; take the per-clip value from the models endpoint or the repo estimators.

import json, os, sys, urllib.request

key = os.environ.get("SUME_API_KEY")
if not key:
    sys.exit("Set SUME_API_KEY first.")

CLIPS = 100
MICROS_PER_CLIP = 625_000  # Wan 3.0, 720p, 5 s, billed
need = CLIPS * MICROS_PER_CLIP

req = urllib.request.Request(
    "https://api.sume.com/v1/balance",
    headers={"Authorization": f"Bearer {key}"},
)
with urllib.request.urlopen(req) as resp:
    balance = json.load(resp)["data"]["balance"]

have = balance["available_amount_usd_micros"]
print(f"have ${have / 1e6:.2f}, need ${need / 1e6:.2f}, state {balance['state']}")
sys.exit(0 if have >= need else 1)

What the check does not cover

• It assumes every job holds at once. If you submit in waves of 20, you only need $12.50 spendable at a time, plus what you spend as waves finish.
• It does not account for retakes. Add a margin for jobs you will re-run.
• A balance read is a snapshot. Other jobs in the same workspace can hold funds between your check and your submit.

Making the script reusable

Pass the per-clip micros and clip count as arguments instead of constants and you can reuse the check across models. Take the per-clip value from the models endpoint when the model has a published pricing line, or from the estimate your submit returns. Keep amounts in integer micros end to end; dividing by a million only for display avoids rounding errors that dollar floats introduce.

Pair the check with an Idempotency-Key on every submit. The docs say not to retry unsafe submit requests without one, and a replay with the same key returns the original job, so a flaky network does not turn into a double charge.

Exit code 1 lets a shell script stop the batch. The error docs say a 402 insufficient_credits is the failure you will otherwise meet at submit time, so the preflight is a courtesy that saves a round trip, not a replacement for handling the 402.