Your China-based supplier has confirmed the product specifications. Your marketing team now needs a set of script options and searchable asset descriptions for several markets. An agency proposes cheaper batch AI processing. Before negotiating the retainer down, establish which work can wait and what happens when a launch-critical item cannot.
The useful buying decision is how to separate scheduled production from exceptions. A discounted model request is one cost component. It does not include every judgment, correction or handoff needed to deliver usable content.
What the published discount actually covers
Checked on October 31, 2026, Anthropic’s Batch processing documentation lists a 50% discount against standard API prices and explains that batches can expire after 24 hours without completing everything. Google’s Gemini Batch API documentation lists 50% of the equivalent model’s standard interactive API cost, with a 24-hour turnaround objective. Neither statement is a finished-content delivery guarantee.
These are terms for the relevant APIs, not a universal discount for AI subscriptions, video generators or agency services. This is an evergreen procurement analysis supported by current documentation, rather than a claim that either feature launched today or that the whole production industry has already cut costs.
Use three work orders before comparing proposals
ONCE proposes a routine batch order, an exception order and an acceptance order. This is a service design method, not a report of measured client results.
The routine order defines work that can wait
Specify the product, market, language, task, approved input revision, submission cutoff and required result time. A candidate pilot might produce script alternatives from confirmed specifications or improve retrieval labels for an existing asset library. Product accuracy and useful labeling still need review.
For a brand sourcing from China, input stability matters as much as time zone differences. If engineering is still changing a performance figure, sending the entire catalog for processing creates avoidable rework. A brief that is waiting for a creative decision is also a poor candidate for volume generation. Settle the decision, or test a small sample first.
The exception order protects time-sensitive work
| Work condition | Suggested route | What to specify |
|---|---|---|
| Stable inputs with waiting time available | Scheduled batch | Cutoff, scope and review slot |
| Fixed launch with too little waiting time | Assess an interactive or manual route separately | Capacity, added cost and feasible delivery |
| Changing facts or unapproved claims | Resolve inputs, or run a small sample | Approval owner and affected items to redo |
An interactive route also needs capacity; it is not automatically a rush guarantee. Agree when progress will be checked, who may authorize additional spending, and when to retry, narrow the release or use manual work. Unattended waiting is not labor, although it can consume the remaining launch window.
The acceptance order separates returned work from usable work
Reconcile each original item with its result, review, correction and accepted delivery. Anthropic distinguishes successful, errored, canceled and expired outcomes, and documents no charge for the latter three described cases. That is its interface rule, not a cross-provider billing rule. An unbilled request can still create coordination work.
Ask for the missing-item list: accepted, needing correction, or awaiting a result. A rerun should identify its scope so already accepted work is not processed again unnecessarily. Technical completion and a correct, publishable script are different acceptance questions.
Calculate savings after operating the service
Net savings for equivalent work = baseline interactive model spend − actual batch model spend − retry or rush costs not already included − added operating costs − the setup allocation for the period.
Hold the task, model, output requirements and acceptance standard constant. Disclose changes instead of attributing them all to batching. Avoid counting the same retry twice. Include other tools, storage and service charges where incurred; do not assume they inherit the model discount. Allocate setup over a defensible workload rather than an imaginary permanently full queue.
Consider a purely hypothetical budget in one currency: baseline model spend of 2,000 units falls to a batch bill of 1,000. Separate retries and rush work cost 400, additional operations cost 300, and allocated setup costs 200. Net savings are just 100 units. These are invented arithmetic inputs, not ONCE rates or a market benchmark. Lower volume could make the allocation larger and the saving negative.
That calculation concerns this processing change. It does not halve filming, art direction, editing or brand review costs. A smaller model bill should not silently reduce the promised creative or language standard.
Write the retainer around routine work and exceptions
Request separate scope for scheduled batches, exception handling and ongoing maintenance. Define the input cutoff for routine work; capacity, escalation and spending approval for exceptions; and ownership of templates, product updates and output checks for maintenance. Fixed fees and usage pricing can both work when the included service is clear.
Pilot a repeatable, non-urgent assignment against a recorded baseline. Keep item-level outcomes, full costs and acceptance times. Expand only when usable output meets the requirement, net savings are positive and exceptions do not displace the normal workload. Without a baseline, establish one before forecasting a percentage saving.
For ONCE’s recurring content services, records of waiting time available, input changes, retry reasons and review effort could improve capacity planning and pricing. That is a proposed use of operating data, not a demonstrated margin gain. Explore more AI content workflow research when defining the next service pilot.
Sources and scope
Primary sources: Anthropic, Claude Platform Docs, “Batch processing”; Google AI for Developers, “Batch API,” marked updated September 17, 2026. Both checked October 31, 2026. Terms apply to the documented services, not automatically to another cloud, intermediary or video tool.
The three work orders, cost model and recommendations are ONCE analysis. Examples are hypothetical; no client project, actual test or achieved saving is claimed. The original AI cover is a conceptual illustration of batch work.
