OpenAI Batch API:
50% Cost Savings — How It Works in 2026

Batch API Pricing vs Standard API

How the Batch API Works

The Batch API processes requests asynchronously — you submit a batch of requests, OpenAI processes them within 24 hours, and you retrieve results when ready:

1. Create a batch file: JSONL file with up to 50,000 requests
2. Upload the file: via Files API
3. Submit the batch: POST to /v1/batches
4. Poll for completion: GET /v1/batches/{batch_id}
5. Retrieve results: Download output JSONL file

Key constraints:

• Maximum 24-hour turnaround (not guaranteed latency)
• Up to 50,000 requests per batch
• Enqueued token limits: 90,000 tokens/minute per model by default
• No streaming — results available only after batch completes

Real-World Batch API Savings

Example: Content Moderation at Scale

Processing 100,000 user reviews per day for sentiment and safety:

• Average: 200 tokens input + 50 tokens output per review
• Daily tokens: 20M input + 5M output
• Standard GPT-4o mini: 20M × $0.15 + 5M × $0.60 = $3.00 + $3.00 = $6.00/day
• Batch GPT-4o mini: 20M × $0.075 + 5M × $0.30 = $1.50 + $1.50 = $3.00/day
• Annual savings: $1,095

Example: Research Paper Analysis Pipeline

Processing 10,000 academic papers monthly with GPT-4o:

• Average: 3,000 tokens input + 800 tokens output
• Monthly tokens: 30M input + 8M output
• Standard: 30M × $2.50 + 8M × $10.00 = $75 + $80 = $155/month
• Batch: 30M × $1.25 + 8M × $5.00 = $37.50 + $40 = $77.50/month
• Annual savings: $930

When to Use Batch vs Real-Time API

Combining Batch API with Prompt Caching

For maximum savings, combine Batch API (50% off) with prompt caching (up to 90% off on cached tokens):

• Long system prompt (10K tokens) shared across all batch requests
• After first request, system prompt cached at 50% discount
• Effective cost on cached prompt tokens in batch: $0.0625/M (GPT-4o mini)
• That's 96% cheaper than standard GPT-4o mini uncached input