LLM Batch API discount comparison
Most major model providers quietly run a second price list: send your prompts through their batch API instead of the synchronous one, and the same model costs half — sometimes 75% less. We compared all 73 batch-priced model pairs on OpenRouter to see who actually discounts the most.
The biggest batch discounts
Top 20 of 73 batch pairs, ranked by input-price discount. Prices are dollars per million tokens, standard vs batch.
| Model | Standard (in / out $/M) | Batch (in / out $/M) | Discount |
|---|---|---|---|
openai/gpt-5.6-sol-pro | $4 / $20 | $1 / $5 | 75% |
deepseek/deepseek-v4.1-flash | $0.3 / $1.2 | $0.112 / $0.336 | 63% |
z-ai/glm-5.3-flash | $0.15 / $0.5 | $0.06 / $0.2 | 60% |
openai/gpt-5-nano | $0.05 / $0.4 | $0.025 / $0.2 | 50% |
openai/gpt-6-luna-pro | $0.1 / $0.5 | $0.05 / $0.25 | 50% |
openai/gpt-6-luna | $0.1 / $0.5 | $0.05 / $0.25 | 50% |
google/gemini-2.5-flash-lite | $0.1 / $0.4 | $0.05 / $0.2 | 50% |
openai/gpt-4.1-nano | $0.1 / $0.4 | $0.05 / $0.2 | 50% |
mistralai/mistral-small-2603 | $0.15 / $0.6 | $0.075 / $0.3 | 50% |
mistralai/ministral-8b-2512 | $0.15 / $0.15 | $0.075 / $0.075 | 50% |
openai/gpt-4o-mini | $0.15 / $0.6 | $0.075 / $0.3 | 50% |
openai/gpt-5.6-luna-pro | $0.2 / $1.2 | $0.1 / $0.6 | 50% |
openai/gpt-5.6-luna | $0.2 / $1.2 | $0.1 / $0.6 | 50% |
openai/gpt-5.4-nano | $0.2 / $1.25 | $0.1 / $0.625 | 50% |
google/gemini-3.1-flash-lite | $0.25 / $1.5 | $0.125 / $0.75 | 50% |
openai/gpt-5-mini | $0.25 / $2 | $0.125 / $1 | 50% |
google/gemini-3.5-flash-lite | $0.3 / $2.5 | $0.15 / $1.25 | 50% |
mistralai/codestral-2508 | $0.3 / $0.9 | $0.15 / $0.45 | 50% |
google/gemini-2.5-flash | $0.3 / $2.5 | $0.15 / $1.25 | 50% |
mistralai/mistral-medium-3.1 | $0.4 / $2 | $0.2 / $1 | 50% |
What is a batch API, exactly?
You upload a file of requests (JSONL), the provider processes it within roughly 24 hours, and you download the results. No streaming, no interactive latency. In exchange for that patience, the same tokens cost a fraction of the synchronous price — typically 50%, and up to 75% on some OpenAI pro-tier models.
The discount applies to both input and output tokens on every pair we checked.
Is batch worth it for your workload?
Batch is a clear win when the task is not interactive:
- Evaluation runs and benchmarks — thousands of prompts, nobody waiting.
- Backfill and data enrichment — labeling, summarizing, classifying archives.
- Scheduled report generation — if the report is due tomorrow morning, a 24-hour window is free money.
- Synthetic data generation for fine-tuning pipelines.
Batch is the wrong tool for chat, agents, or anything with a user in the loop. If you can only batch half your traffic, batching that half still cuts your blended cost by roughly 25%.
Worked example
Processing 10M input + 2M output tokens per day on gpt-4o-mini:
- Synchronous: (10 × $0.15) + (2 × $0.60) = $2.70/day → $81/month
- Batch: (10 × $0.075) + (2 × $0.30) = $1.35/day → $40.50/month
Same model, same outputs, half the bill. Run your numbers in the batch API savings calculator.
Which providers offer batch pricing?
In the live OpenRouter catalog: anthropic, deepseek, google, mistralai, moonshotai, openai, x-ai, z-ai. The dominant pattern is exactly 50% off; OpenAI's pro reasoning tier discounts up to 75%, and DeepSeek's V4.1 Flash reaches ~63% off input.
FAQ
How much does batch API save?
The median across 73 model pairs is 50% off both input and output tokens. The range runs from about 40% to 75%.
How long does batch processing take?
Providers target completion within 24 hours; many batches finish in a few hours. It is not a guaranteed instant queue.
Can I batch chat traffic?
No — batch is for offline workloads. Anything with a user waiting on the response belongs on the synchronous API.
Is output quality the same in batch mode?
Yes. It is the same model with the same weights; batch only changes when the request is processed and what it costs.
Check any model's standard vs batch rate in the OpenRouter pricing calculator — 434 models, live rates. More cost questions answered in the answers index.