Batch API shipping

Half-price, no rush.

Some workloads don't need answers in real time — overnight backfills, classification of months of records, summary generation across an archive. The Batch API lets you submit a list of message-create requests, walk away, and pick them up when the batch finishes. Cost is about 50% of synchronous calls.

The decision rule.

Requests in, results out.

1. Build a JSONL file, one request per line, each with a custom_id you'll use to match results back.
2. Submit the batch. The API returns a batch_id.
3. Poll status until it's ended.
4. Download results. A JSONL file in the same order, each row keyed by your custom_id.

What batch is great for.

• Tagging an archive. 50,000 articles, classify each into one of 8 categories.
• Backfill summaries. Five years of meeting notes, summarize each.
• A/B prompt experiments. Run the same 1,000 inputs through 3 prompt variants; compare outcomes.
• Eval datasets. Score model responses against a rubric.

Don't lose a batch.

• Persist the batch_id and the input JSONL the moment you submit. Crashes happen.
• Use stable custom_ids so you can rejoin to source records.
• Cap each batch around 10,000–50,000 requests. Smaller chunks recover faster from a partial failure.
• Track cost per batch. Log input tokens, output tokens, completed-vs-failed counts.
• Combine with caching when prefixes repeat across requests (same system prompt, same docs).

Related products.

Every Anthropic surface has its own page on this guide.