Batch AI Processing Cost — 1M Documents/Month with Batch API Discounts

Scenario

An offline pipeline processes 1M documents per month — classifying, summarizing, or extracting structured data. The work is not latency-sensitive (overnight batch is fine) so we can use the batch APIs offered by OpenAI, Anthropic, and others for 50% off. Each document is ~2,000 input tokens and the model produces ~200 tokens of output.

Note: the cost numbers below show ON-DEMAND pricing — apply 50% discount manually for batch API on OpenAI / Anthropic / Google. The calculator does not yet auto-apply batch pricing.

Monthly cost across recommended models

Calculated at 2000M input tokens + 200.0M output tokens, with 5% prompt cache hit rate.

💡 Switching from Claude Haiku 4 5 to Deepseek Chat saves $2291/month (79% reduction).

Why these models

Batch processing rewards the cheapest reliable model. Quality matters less per doc since you can re-process failures. GPT-5 Mini and Claude Haiku 4.5 both offer 50% off via batch API. DeepSeek is the absolute cheapest but lacks a formal batch discount (their on-demand price is already lower than competitors batch prices). Gemini Flash with batch API is competitive too.

Key insights

• 1. Use batch APIs whenever you can — 50% discount is the easiest win in LLM cost optimization.
• 2. For 1M+ docs/month, even a 10% prompt template optimization saves hundreds of dollars. A/B test prompts on small batches first.
• 3. Failures matter: if 2% of docs fail and need retry on a more expensive model, the effective cost rises. Track failure rates per model.
• 4. Self-hosting an open model on GPU rentals (Modal, Runpod) can beat batch APIs at 10M+ docs/month — worth evaluating at high volume.

Cost at different scales