Batch Saves Money, but a Cheaper Model at Equal Quality Saves More
Batch Saves Money, but a Cheaper Model at Equal Quality Saves More
Batch serving costs less than sync for the same model on the same work, with no change in reported score. But on tasks where several models tie on quality, the choice of model moved cost more than the choice of mode: a cheaper model at its sync price undercut a pricier one in batch.
Batch is a discount, not a quality trade-off
That parity comes from paired measurements: the same model, the same work, both modes. In these results, sync buys responsiveness and nothing else. Some results exist in one mode only, and there the accuracy cost of a batch saving is unknown.
The stated discounts run from 16% to 77% and mostly sit near half [^1][^2][^3][^4]. Claude Opus 5.5 writing short promotional messages scores 8.71 out of 10 (±0.21) in either mode, at $0.008 in batch against $0.02 in sync [^5]. Extracting structured fields, it scores 9.93 (±0.16) at $0.07 against $0.13 [^6].
The spread is wide, and it depends on both the model and the task. A discount seen on one kind of work should not be assumed for another.
A cheaper tied model beats a cheaper mode
Halving a price is useful. Switching to a model that ties on score for a fraction of the price is often worth more. On three tasks, a cheaper model in sync undercuts a pricier model in batch, with the two scores tied within their intervals:
| Task | Pricier model, in batch | Cheaper model, in sync |
|---|---|---|
| Generating visual theme configurations | Claude Opus 5.5: 8.7 (±0.24), $0.04 [^7] | Claude Haiku 5.5: 8.43 (±0.2), $0.0023 [^7] |
| Triaging engagement opportunities | Claude Sonnet 5.5: 8.5 (±0.29), $0.0024 [^8] | Claude Haiku 5.5: 8.54 (±0.3), $0.0008 [^8] |
| Checking policy eligibility for a persona | Gemini 3.8 Flash: 8.66 (±0.2), $0.0026 [^9] | DeepSeek V4.1 Flash: 8.95 (±0.22), $0.0007 [^9] |
On visual theme configurations, Claude Opus 5.5 costs more than ten times as much as Claude Haiku 5.5, even with Opus on its batch discount and Haiku on sync pricing [^7].
Mode still matters when models do not tie. On writing section prompts, Claude Opus 5.5 scores 9.11 (±0.22) at $0.05 in batch, while Claude Sonnet 5.5 scores 8.17 (±0.46) at $0.05 in sync [^10]. The intervals do not overlap, so a team that can accept batch gets the stronger model for the weaker model’s sync price.
How to route with this
The evidence suggests an order of decisions:
- Compare models in the same mode where you can. Price depends on mode, so setting one model’s batch price against another’s sync price can make the first look cheaper than it is.
- Pick the cheapest model whose score clears your quality bar.
- Then move that model to batch where the work can wait.
The last step depends on the model, not the pipeline. Batching is a per-model capability, and a model with no batch price for a task is costed at its sync price inside a batch-when-possible pipeline. DeepSeek V4.1 Flash, the cheap choice on policy eligibility, appears in these findings only with a sync price [^9]. A missing batch price is not proof that a model lacks a batch lane, but routing its jobs to batch saves nothing unless a lower batch price exists for that task.
What the data cannot tell you about waiting
The benchmark reports no batch turnaround time. Every latency figure in this cycle is for synchronous calls under a 600-second per-call deadline. Whether a job can wait for batch is a judgement the data cannot make for you. If a person or a downstream step is waiting, keep it in sync and size the response-time budget on the 90th percentile rather than the median.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your system needs, see which models tie at that bar, and compare their batch and sync prices on your task before committing a pipeline to either mode.