The Batch API, Fully Specified
Batch API — published mechanics
| Discount vs Standard | 50% off the Standard input/output rate |
|---|---|
| Completion window | Fixed at 24h — the only accepted value; batches often finish sooner |
| Rate-limit pool | Separate from standard RPM/TPM — Batch never consumes them |
| Output tokens in the rate limit | Not counted — queue limits count queued prompt tokens only (each response is still capped by the model's max output) |
| Requests per batch | Up to 50,000 |
| Input file size | Up to 200 MB |
| Batch creation rate | Up to 2,000 batches per hour |
| On expiry | Unfinished requests are cancelled; completed ones are returned and billed |
| Platforms named as lacking Batch | None, as of the latest check — see the note below the table |
Standard vs Batch price, every model that offers Batch
| GPT-6 Astra | Standard $10.00 → Batch $5.00 /MTok in · $50.00 → $25.00 out |
|---|---|
| GPT-6.1 Sol | Standard $2.00 → Batch $1.00 /MTok in · $10.00 → $5.00 out |
| GPT-6.1 Terra | Standard $1.00 → Batch $0.50 /MTok in · $5.00 → $2.50 out |
| GPT-6 Sol | Standard $2.00 → Batch $1.00 /MTok in · $10.00 → $5.00 out |
| GPT-6 Luna | Standard $0.10 → Batch $0.05 /MTok in · $0.50 → $0.25 out |
| GPT-5.6 Sol | Standard $4.00 → Batch $2.00 /MTok in · $20.00 → $10.00 out |
| GPT-5.6 Terra | Standard $2.00 → Batch $1.00 /MTok in · $12.00 → $6.00 out |
| GPT-5.6 Luna | Standard $0.20 → Batch $0.10 /MTok in · $1.20 → $0.60 out |
| GPT-5.5 | Standard $5.00 → Batch $2.50 /MTok in · $30.00 → $15.00 out |
| GPT-5.5 Pro | Standard $30.00 → Batch $15.00 /MTok in · $180.00 → $90.00 out |
| GPT-5.4 | Standard $2.50 → Batch $1.25 /MTok in · $15.00 → $7.50 out |
| GPT-5.4 Mini | Standard $0.75 → Batch $0.375 /MTok in · $4.50 → $2.25 out |
| GPT-5.4 Nano | Standard $0.20 → Batch $0.10 /MTok in · $1.25 → $0.625 out |
| GPT-5.4 Pro | Standard $30.00 → Batch $15.00 /MTok in · $180.00 → $90.00 out |
The Batch API isn't a discount code for the normal endpoint — it's a separate way of submitting work, with its own guarantees, its own rate-limit pool and its own trade-offs against a synchronous request.
What you're trading
The discount — 50% off Standard — applies uniformly to every model that offers Batch, which makes it one of the simplest levers on this site: if a workload can wait, Batch is close to strictly cheaper for the same tokens. What you give up is timing. The completion window is fixed at 24 hours; it is the only value the API accepts, and there's no way to pay for a shorter one. Batches often finish sooner, but nothing promises it.
A separate pool, on purpose
Batch requests don't draw down your per-model RPM or TPM — they run against their own limits: a per-batch cap on requests and file size, a per-model ceiling on queued prompt tokens, and a cap on how many batches you can create per hour. A large background job therefore can't starve your live traffic of headroom.
What "no output-token limit" means
OpenAI's guide says the Batch API "currently has no output-token limit." This page used to read that as "a Batch response can run longer than a normal one." It can't: the sentence sits in the guide's rate-limit section and means the queue counts queued prompt tokens, not output. Each response is still capped by the model's own maximum output.
When a batch expires
If a batch doesn't finish in its window, unfinished requests are canceled, and completed ones are returned — and billed.
Platforms
This page used to list Amazon Bedrock, Google Vertex AI and Microsoft Foundry as platforms where Batch wasn't offered. No OpenAI page checked in October names any platform that way, so check your provider's own documentation before designing a Bedrock or Azure deployment around Batch.
The tables above give every published limit and price every model that offers Batch.
Verified 2026-10-28 against https://developers.openai.com/api/docs/guides/batch.
Could not confirm: Which cloud platforms lack Batch. Until 2026-10-01 this site listed Amazon Bedrock, Google Vertex AI and Microsoft Foundry; none of the OpenAI pages checked on 2026-10-01 names any platform as lacking Batch, and the Bedrock guide defers service tiers to AWS's own pricing page. So no platform is listed — check your provider's docs.
Checked: https://developers.openai.com/api/docs/guides/batch · https://developers.openai.com/api/docs/pricing · https://developers.openai.com/api/docs/guides/amazon-bedrock