Batch Is Cheaper — Except When It Isn't
The batch discount is real, but check it task by task
Where the Fronset benchmark measured the same model on the same work in both batch and synchronous mode, batch cost less per task run in nearly every pair, and the quality score did not change. But the size of the saving swings widely by model and task, and in two pairs batch cost more. The batch price has to be checked for each model and task, not assumed.
Where the discount holds
Batch mode means submitting work to be returned later rather than waiting on each call. The cleanest evidence is a paired measurement: one model, one task, both modes, with an identical score out of 10 in each, so the cost difference comes from the mode and not from a change in quality. Claude Opus 5.5 classifying whether claims continue one another scores 8.64 (±0.24) at $0.19 per task run in batch against $0.36 synchronously [^1]. Claude Haiku 5.5 validating research queries scores 8.43 (±0.47) at $0.0002 against $0.0004 [^2], and Gemini 3.8 Flash rewriting a metadata paragraph scores 7.64 (±0.22) at $0.0031 against $0.005 [^3].
Pairs with scores like these exist only for some tasks. For most others only cost is reported, so a batch score cannot be assumed for them [^4].
How much it varies
The discount is not a fixed rate. Claude Opus 5.5 is stated at 26% less on scoring generated content for relevance, at $0.06 against $0.08 [^5], but 91% less on scoring a content set for relevance, at $0.04 against $0.46 [^6]. GPT-6 Luna is stated at 85% less on structured content summarization, at $0.0004 against $0.0028 [^7].
At very small prices the percentage stops mattering. Gemini 3.8 Flash on language identification costs $0.0004 in batch against $0.0005 synchronously [^8]: the percentage gap is real, but the dollar gap is negligible.
Where it reverses or cannot be assumed
Three paired measurements run against the pattern, each specific to one model and one task. The same GPT-6 Luna scores 8.42 (±0.46) in both modes on scoring generated content for relevance, yet batch costs $0.0038 against $0.0022 synchronously [^5]. Gemini 3.5 Flash Lite on the same kind of scoring costs $0.0061 in batch against $0.0038 [^5], and Xiaomi MiMo V2.6 Flash costs $0.0001 per task in both modes on engagement opportunity triage [^9].
The discount also depends on a batch lane existing at all. Batching is a per-model capability: a model without it is priced at its sync cost inside a batch-when-possible pipeline, and the benchmark does not state which models lack one. The absence of a batch price for a model is not evidence that it lacks one, and a configuration measured in one mode only, such as Gemini 3.8 Flash on engagement opportunity triage in batch [^9], supports no inferred saving in either direction.
When to batch and when not to
Make batch the default for deferrable work when the model has a measured batch lane on that specific task, its batch price is below its sync price, and the paired score clears the quality bar your system needs. Use sync where the batch price is at or above the sync price, as for GPT-6 Luna and Gemini 3.5 Flash Lite on relevance scoring [^5], and for anything a person is waiting on.
Price the model before the mode. On short promotional campaigns, Claude Haiku 5.5 scores 7.45 (±0.38) and Claude Sonnet 5.5 scores 7.78 (±0.31), a tie because the ranges overlap, yet Haiku synchronously costs $0.0034 while Sonnet in batch costs $0.03 [^10]. Batch makes a given model cheaper; it does not make an expensive model competitive with a cheaper one that clears the same bar.
Budget for the wait. Half of all batches return within 5.0 minutes and 79.2% within 15 minutes, but 5.3% of completed batches took longer than an hour [^11]. A one-hour deadline cancels batches that run past it and redispatches them to a different model, which happened to 3.9% of submitted batches [^11]. A redispatch can change the output as well as the delay, so plan for it wherever a different model’s answer would be a problem.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your system needs and compare batch against sync pricing for the models and tasks you are considering. Check the paired figures for your own configuration before moving production traffic.