What the Batch API is
OpenAI's Batch API (and the related Flex processing tier) lets you submit requests that do not need an immediate answer. Results come back within 24 hours instead of seconds, and in exchange every token — input and output — costs 50% of the standard rate. There is no quota application and no minimum volume; the only price is latency.
Verified batch rates: GPT-6 family
| Model | Standard in/out | Batch in/out | You save |
|---|---|---|---|
| GPT-6 AstraOpenAI API pricing, Oct 2026 | $10.00 / $50.00 | $5.00 / $25.00 | 50% on every token |
| GPT-6.1 SolOpenAI API pricing, Oct 2026 | $2.00 / $10.00 | $1.00 / $5.00 | 50% on every token |
| GPT-6 LunaOpenAI API pricing, Oct 2026 | $0.10 / $0.50 | $0.05 / $0.25 | 50% on every token |
The discount applies to the whole request — input, output, and (on OpenAI) it stacks with prompt caching discounts. A batched, cached Sol input token can cost a fraction of a cent per thousand.
Which providers offer batch pricing?
- OpenAI — yes. Batch API at 50% off across the GPT-6 family, plus Flex processing at batch rates for lower-priority interactive traffic.
- Anthropic — yes, on select models. A 50% batch discount is offered on parts of the Claude lineup (Anthropic's pricing documentation lists Claude Haiku 4.5 at $0.50/$2.50 batch) — confirm the current list on Anthropic's pricing page before planning around it.
- xAI — not currently. Grok 4.7's rate card shows no batch discount; the lever there is the cached-input rate and staying under the 200K-token band.
- DeepSeek — not needed. Off-peak rates already halve peak pricing, and automatic caching discounts input to $0.003/1M on Flash — batching would add little.
When batching is the right call
Anything where nobody is waiting on the response: nightly report generation, dataset labeling, bulk classification, content moderation queues, embedding jobs (text-embedding-3-small costs $0.01/1M on batch, per OpenAI's pricing page), evaluation harnesses, and backfill pipelines. If a job runs on a cron schedule, it is a batch candidate by definition.
Standard: 10 × $2 + 5 × $10 = $70 · Batch: 10 × $1 + 5 × $5 = $35
That $35 saving required zero code changes to prompts or models — just a different endpoint. For recurring nightly jobs, batching is usually the first discount to capture because it needs no prompt engineering and never changes answer quality.
Stacking the discounts
The biggest bills get cut by combining levers, not picking one: route to the cheapest capable tier first, batch everything that can wait, and cache repeated prefixes. A batched Luna job with a cached system prompt can run at roughly 5% of the equivalent interactive Astra cost. Work the math for your own mix with our bill estimator.
Quick answers
How much does the OpenAI Batch API save?
50% on both input and output tokens versus standard pricing. Verified October 2026: GPT-6 Astra drops to $5/$25, GPT-6.1 Sol to $1/$5, and GPT-6 Luna to $0.05/$0.25 per million tokens. Jobs are returned within 24 hours, so the discount rewards workloads that can wait.
Does the Batch API change answer quality?
No. The Batch API runs the same models with the same weights and the same decoding — the only differences are turnaround time (up to 24 hours) and scheduling priority. Answers are identical to standard API calls, which is why batch is the default choice for offline evaluation, labeling, and report workloads.
Can I combine batch pricing with prompt caching?
On OpenAI, yes — the discounts stack. Cached input tokens inside a batch job are priced at the cache rate applied to the already-discounted batch input rate, which makes repeated-prompt offline workloads the cheapest way to run on the rate card. The same stacking logic applies on Google's Batch API.
Is there a minimum volume for batch pricing?
No. There is no minimum spend, request count, or commitment — the 50% discount applies per request from the first call. The only tradeoff is latency: results can take up to 24 hours. That makes batch suitable for scheduled and offline workloads of any size, from nightly reports to one-off backfills.