Rate
Batch API: when half price is worth the wait
Most AI providers will sell you the same model for half price if you stop asking for an instant answer. That is the deal behind the Batch API: you upload a file of requests, the provider works through it when it has spare capacity, and you collect the results later. This guide puts dollar figures on that discount for the current models on our prices page, using the official list prices on our prices page, and shows where it is worth the wait and where it is not.
Who offers what
| Provider | Batch discount | Notes |
|---|---|---|
| Anthropic (Claude) | 50% off input and output | All current Claude models |
| OpenAI (GPT-6, GPT-5.6) | 50% off input and output | OpenAI says results come back within 24 hours. GPT-6.1 Terra also lists Flex at the same 50% off |
| Google (Gemini) | 50% off input and output | Google also lists Flex processing for Gemini 4 Argon at the Batch price |
| Mistral | 50% off | Listed on Mistral’s API pricing docs as a general batch discount |
| xAI (Grok 4.7) | 20% off | $1.60 input, $0.40 cached input, $4.80 output per million |
| DeepSeek | No batch discount listed | Off-peak hours are half price instead (see below) |
Two examples from the official pages: GPT-6.1 Terra drops from $1 / $5 to $0.50 / $2.50 per million tokens through Batch or Flex, and Gemini 4 Argon drops from $2 / $10 to $1 / $5 (until its introductory price ends on November 30).
What 100,000 summaries cost, standard vs batch
A typical batch job: summarise 100,000 documents of 10,000 tokens each, with a 300-token summary per document. That is 1 billion input tokens and 30 million output tokens. The formula is the same as in our cost per 1,000 requests guide, just with a bigger number of requests. Sorted by standard price, cheapest first.
| Model | Price in / out (per 1M) | Standard | Batch | You save |
|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 | $115.00 | $57.50 | $57.50 |
| GPT-6 Luna | $0.10 / $0.50 | $115.00 | $57.50 | $57.50 |
| GPT-5.6 Luna | $0.20 / $1.20 | $236.00 | $118.00 | $118.00 |
| DeepSeek V4.1 Flash* | $0.30 / $1.20 | $336.00 | $168.00 off-peak | $168.00 |
| Gemini 3.5 Flash-Lite | $0.30 / $2.50 | $375.00 | $187.50 | $187.50 |
| Mistral Large 3§ | $0.50 / $1.50 | $545.00 | $272.50 | $272.50 |
| Mistral Large 4§ | $0.70 / $2.10 | $763.00 | $381.50 | $381.50 |
| Gemini 3.8 Flash† | $0.75 / $3.75 | $862.50 | $431.25 | $431.25 |
| GPT-6.1 Terra | $1 / $5 | $1,150.00 | $575.00 | $575.00 |
| DeepSeek V4 Pro* | $1.32 / $3.96 | $1,438.80 | $719.40 off-peak | $719.40 |
| Grok 4.7 | $2 / $6 | $2,180.00 | $1,744.00 | $436.00 |
| Claude Sonnet 5.5 | $2 / $10 | $2,300.00 | $1,150.00 | $1,150.00 |
| Gemini 4 Argon† | $2 / $10 | $2,300.00 | $1,150.00 | $1,150.00 |
| GPT-6.1 Sol | $2 / $10 | $2,300.00 | $1,150.00 | $1,150.00 |
| Gemini 3.1 Pro | $2 / $12 | $2,360.00 | $1,180.00 | $1,180.00 |
| Claude Opus 5.5 | $4 / $20 | $4,600.00 | $2,300.00 | $2,300.00 |
| GPT-5.6 Sol‡ | $5 / $30 | $5,900.00 | $2,950.00 | $2,950.00 |
| Claude Fable 5.1 | $10 / $50 | $11,500.00 | $5,750.00 | $5,750.00 |
| GPT-6 Astra | $10 / $50 | $11,500.00 | $5,750.00 | $5,750.00 |
* DeepSeek prices are peak prices. DeepSeek lists no Batch API discount; the “batch” column shows the same job run entirely in off-peak hours, which DeepSeek bills at half price.
† Promotional prices. Gemini 4 Argon becomes $4 / $20 from December 1, 2026, which makes this job $4,600 standard or $2,300 via batch. Gemini 3.8 Flash becomes $1.50 / $7.50 from January 1, 2027: $1,725 standard or $862.50 via batch.
‡ $5 / $30 is the list price OpenAI’s main pricing page showed before it stopped listing per-token prices. OpenAI’s developer pricing page and model page show a promotional $4 / $20 until November 30, 2026 (this job: $4,600 standard, $2,300 via batch).
§ Assumes Mistral’s general 50% batch discount applies to these models; Mistral Large 4 is still in preview, so check Mistral’s docs before relying on it.
Three things the table shows
1. Batch can reorder the ranking. At standard prices, Grok 4.7 ($2,180) is slightly cheaper for this job than Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon ($2,300 each). Through the Batch API it flips: those three drop to $1,150, while Grok 4.7’s 20% discount only takes it to $1,744. If your job can wait, compare batch prices, not list prices.
2. A cheaper model beats a discount. Batch halves the bill, but switching from Claude Sonnet 5.5 to GPT-6.1 Terra halves it too, and doing both cuts the job from $2,300 to $575. Moving down to Claude Haiku 5.5 or GPT-6 Luna via batch brings it to $57.50. Test whether the smaller model is good enough for the task before paying for a bigger one at half price.
3. DeepSeek’s off-peak hours act as a batch discount without the queue. DeepSeek’s peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; at all other times, including weekends and Chinese public holidays, it charges half. A job started outside those hours costs the same as a 50% batch discount, and the answers still come back in real time. DeepSeek V4 Pro off-peak ($719.40) costs more than GPT-6.1 Terra via batch ($575.00) for this job, while DeepSeek V4.1 Flash off-peak ($168.00) lands between GPT-5.6 Luna and Gemini 3.5 Flash-Lite via batch.
When batch is worth the wait
Batch suits work where nobody is waiting on the other end:
- Back-office processing: summarising, tagging or translating a document archive, classifying support tickets, extracting fields from invoices.
- Evaluations: running a test set through several models to compare them. Many runs at half price add up quickly.
- Content backfills: generating product descriptions, alt text or metadata for a whole catalogue at once.
- Nightly jobs: anything you already run on a schedule, where results are needed tomorrow morning rather than now.
It does not suit chat, coding assistants, agents or anything else where a person or a program is waiting for the next answer. An agent that has to call the model 30 times in a row cannot wait hours between steps, which is why prompt caching matters more for agents than batch does.
How to decide in practice
- Price the job at list price first. Use the cost calculator with your own token counts. If the whole job costs a few dollars, the discount may not be worth changing your code.
- Turn on the batch setting. The calculator has a “Use the Batch API” option that applies the 50% discount to the models where we record it, so you can see whether the ranking changes. It does not yet apply the batch discount for Mistral Large 4, xAI’s smaller 20% discount or DeepSeek’s off-peak price; use the figures in this guide for those.
- Check promo end dates. A batch job priced on Gemini 4 Argon’s introductory price will cost twice as much after November 30. Our prices page lists each model’s current price and any promo note.
- Compare two candidates side by side on the compare pages, for example Claude Sonnet 5.5 vs GPT-6.1 Sol.
For the basics behind these numbers (tokens, input vs output, caching), see AI API pricing explained. Prices change often; we recheck each provider’s official pricing page regularly and update this guide when they change.