Batch APIs discount 50%: what a Sume bulk run changes instead

Does a Sume bulk run get a batch discount? I found none in the docs. OpenAI's Batch guide, Anthropic's batch page and the Gemini Batch API guide (all read 2026-10-04) each state a 50% discount in exchange for asynchronous turnaround. The Sume Bulk runs docs describe no batch rate: every child is an ordinary Format run, admitted through the same wallet, workspace concurrency and spend-cap checks as a single call.

That is a statement about the documentation, not about future pricing. If you need a number, read your run receipts and the billing record, not a blog post.

What a queue gives you instead

A queue gives control, not a price break. You choose concurrency from 1 to 16, the server keeps that many children in flight and refills slots as runs finish, and you set generation_spend_cap_usd per item, up to 500. The Bulk runs docs list no queue-level cap, so your worst case is the number of items times the per-item cap, and you should set a cap deliberately.

Each completed run's usage.billable_amount_usd_micros carries the generation spend attributed to it, and usage is null when it could not be read. GET /v1/usage stays the authoritative billing record.

Reading your own cost

Do not compare a discount with a run cost from memory. Run ten items, read each receipt's usage.billable_amount_usd_micros, and add them. That figure excludes the LLM turn, and the docs say debited_usd_micros is the real cost, so read both before you quote a per-video number to a finance team. usage is null, not zero, when it cannot be read; treat null as unknown.

Then compare like with like. A 50% discount on a text step that costs cents saves cents; the video step is where the spend is, and that step is outside what any of the three text batch pages describes.

Discount versus latency

Worst-case spend before you press send

Compute the ceiling for a queue from your own caps. The numbers below are examples, not prices.

items = 100
per_item_cap_usd = 12.50     # your choice, must be 0 < cap <= 500
assert 0 < per_item_cap_usd <= 500
worst_case = items * per_item_cap_usd
print(f"worst case for {items} items: ${worst_case:,.2f}")
for c in (4, 16):
    print(f"concurrency {c}: at most {c} items in flight, "
          f"${c * per_item_cap_usd:,.2f} exposed at once")

When to use which

Use a vendor batch for text where a day is fine and the discount matters. Use a queue when each result is a media file you want back item by item with a per-run webhook, and when you want to set how many run at once. The two do not compete: copy in a batch, video in a queue is a sensible split.