Batch API Discounts & Cache Stacking Rules Across 8 Providers
Batch APIs promise 50% off asynchronous workloads, but real-world billing diverges sharply: Together AI discounts 0% on serverless chat, Bedrock excludes Claude 5.x and gives 25% on Mistral Large 2, and Gemini 3.1 Pro Preview keeps cached reads at full standard price.
Three Distinct Batch + Cache Regimes
1. Multiplicative Stacking (OpenAI GPT-5+, Anthropic Claude, Gemini 4 Carbon & 3.8 Flash): Batch cuts input, output, AND cached-read rates by 50% (plus cache writes on OpenAI and Anthropic). For example, Claude Haiku 5.5 Batch drops from $0.10/$0.50/$0.01 to $0.05/$0.25/$0.005 (<=100K tier).
2. Batch on Uncached Tokens Only, Cache at Standard Rate (Gemini 3.1 Pro Preview): Google's Batch API documentation explicitly bills Gemini 3.1 Pro Preview cached reads ($0.20 <=200K, $0.40 >200K) and explicit cache storage ($4.50/1M/hr) at 'Same as Standard', while discounting uncached input and output by 50% ($1.00/$6.00 <=200K).
3. Batch Disables Caching Entirely (Groq & Amazon Bedrock): Groq's 50% Batch (24h–7d window on 5 text models) does not stack with prompt caching. Amazon Bedrock explicitly restricts prompt caching to on-demand inference endpoints—never Batch.
The Four Biggest Batch Pricing Traps
Together AI (0% Serverless Chat Batch Discount): Although Together AI's marketing mentions up to 50% lower costs, its Batch API documentation specifies that only openai/whisper-large-v3 receives 50% off—all serverless chat models run at standard 100% rates in a separate rate-limit pool.
Amazon Bedrock Claude 5.x (N/A) & Mistral Large 2 (25% off): Bedrock lists Claude Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5 as N/A for Batch inference (use Claude Sonnet 4.6 at $1.50/$7.50 or Opus 5 at $2.50/$12.50 for 50% Bedrock Batch). Meanwhile, Bedrock's Mistral Large 2 (24.07) drops from $2.00/$6.00 to $1.50/$4.50 in Batch—a 25% discount, not 50%.
DeepSeek (No Batch API—Use Off-Peak Instead): DeepSeek has no Batch API; instead, it automatically applies 50% Off-Peak pricing outside Peak UTC windows (01:00–04:00 and 06:00–10:00 UTC Monday–Friday, excluding Chinese public holidays).
Test These Numbers on Your Own Workload
All calculations execute locally in your browser with zero server calls.