Batch API Pricing: Which Providers Actually Discount It (2026)
If your workload can tolerate a delay — overnight processing, bulk classification, evaluation runs, offline document analysis — batch APIs are the single easiest discount to claim, because you don't have to change your prompts or your model choice to get it. But "batch is discounted" isn't universal, and the exact discount isn't always what you'd assume. Here's what each provider we track actually publishes, as of August 2026.
What a batch API actually is
Instead of sending requests one at a time and waiting on each response, you submit a set of requests together; the provider processes them asynchronously (typically with a completion window measured in hours, not seconds) and charges a lower per-token rate in exchange for not needing to serve you in real time. The trade is latency for cost — batch is a poor fit for anything user-facing and a strong fit for anything that runs in the background.
The discount, provider by provider
| Provider | Batch discount | Notes |
|---|---|---|
| OpenAI | 50% off input and output | Same rate across all current models |
| Anthropic | 50% off input and output | Same rate across all current models |
| Google (Gemini) | 50% off input and output | Same rate across all current models |
| Mistral | 50% off input and output | Covers chat completions, embeddings, and OCR |
| xAI (Grok) | 50% off input, output, cached, and reasoning tokens | Image/video generation billed at standard rate even via batch |
| Qwen (Alibaba, DashScope) | 50% off input and output | Cannot be combined with prompt caching — see below |
| Moonshot (Kimi) | 40% off (pays 60% of standard rate) | Published for Kimi K2.5, K2.6, and K2.7-Code only — no published batch rate for K3 at time of writing |
| DeepSeek | No published batch discount | Batch-style requests bill at the standard per-token rate |
| Z.AI | No published batch discount found | Check Z.AI's own docs directly — this may exist without being publicly documented |
Five of the nine providers we track (OpenAI, Anthropic, Google, Mistral, xAI) publish a flat, uniform 50% batch discount. Two more (Qwen, Moonshot) offer a discount with a real caveat attached. Two (DeepSeek, Z.AI) don't appear to offer one at all — DeepSeek's own pricing page is explicit about this rather than silent on it.
Two nuances almost every generic guide skips
Qwen: batch and prompt caching are mutually exclusive. DashScope's own documentation states that if a model supports both batch calls and context caching, you have to pick one — you can't stack a caching discount on top of the batch discount. If your workload has a large repeated prefix and can tolerate batch latency, run the numbers both ways before assuming caching automatically wins; for very high-volume repeated-prefix workloads, the flat 50% batch rate can beat a partial cache-hit rate. (Compare with the cache savings calculator, which models the caching side.)
Moonshot's discount doesn't cover their newest model. K2.5, K2.6, and K2.7-Code have a published batch rate; K3 — the model most people reach for today — doesn't yet. If you're planning around Kimi K3 specifically, don't assume the 40% figure applies until Moonshot publishes a rate for it.
Should you combine batch with caching?
Where a provider allows it (unlike Qwen), stacking a 50% batch discount with a caching discount on top compounds: if a model's cached-read rate is already 90% off standard, and batch adds another 50% off whatever rate applies, the repeated portion of a batch job can end up costing a small fraction of standard real-time pricing. The catch is that not every provider documents how caching and batch interact — some apply the batch discount to the already-cached rate, others may not support caching within the batch endpoint at all. Treat any stacked estimate as optimistic until you've confirmed the actual behavior in that provider's current docs.
Practical takeaway
If part of your workload doesn't need a synchronous response — nightly summarization, dataset labeling, backfilling embeddings, running an eval suite — check whether your model's provider is one of the five with a flat 50% batch discount before you build around real-time calls by default. Batch pricing, unlike per-token list prices, doesn't move often once published, but it isn't universal and it isn't always the flat 50% figure people assume — verify the current rate on the provider's own pricing page before committing a production workload to it, since we don't carry batch rates in our live pricing pipeline (see methodology for what we do track automatically).