Batch Processing: Automating Many Similar Tasks

The idea

Some tasks are really many identical tasks: migrate 500 files, classify 2,000 texts, translate every product description. You don't need to chat 500 times for that – batch processing handles them automatically.

Path 1: headless fan-out

Claude Code also runs without an interactive session via claude -p "task". Inside a script loop over a file list, this becomes a fan-out: one invocation per file. The official docs recommend three steps: generate a task list, build the loop, test on 2–3 files first and sharpen the prompt – then run it on everything. --allowedTools limits what the agent may do unattended.

Path 2: provider batch APIs

For pure model requests without tools there are batch APIs: you submit many requests together and they are processed asynchronously. Anthropic (Message Batches API) and OpenAI (Batch API) charge 50% of the standard prices for this (as of July 2026). At Anthropic, most batches finish in under an hour; at OpenAI a 24-hour window applies (as of July 2026).

Batch or interactive?

Batch pays off when the tasks are independent and uniform and the result is not needed immediately. Stay interactive when you want to steer, review, or correct along the way – or when the task only becomes clear while working.