	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Anthropic Claude API > Claude Batch API

Claude Batch API

Author: Venkata Sudhakar

The Claude Batch API is designed for large-scale asynchronous processing where you have hundreds or thousands of items to process and do not need immediate responses. Instead of sending 10,000 invoice classifications one by one at real-time rates, you submit them all in one batch request, Claude processes them over up to 24 hours, and you retrieve results when complete. The cost is 50% lower than the standard API for the same model. This makes tasks that would be prohibitively expensive or rate-limited at real-time pricing perfectly affordable as nightly batch jobs.

A batch is a list of requests, each with a custom_id you assign for tracking and a standard messages API params object. Submit with client.beta.messages.batches.create(), receive a batch_id, then poll until processing_status is "ended". Results are matched back to your custom_ids. The Batch API supports all Claude models, system prompts, tools, and max_tokens - it is the exact same API surface as real-time, just asynchronous and half the price. Ideal for overnight ETL jobs, bulk document classification, and large-scale data enrichment.

The below example shows an accounts payable team processing supplier invoices overnight - classifying by department, extracting amounts, and flagging invoices needing manager approval - all in one batch at half the real-time cost.

import anthropic, json, time

client = anthropic.Anthropic(api_key="your-api-key")

# Supplier invoices (in production: fetched from your ERP system)
invoices = [
    {"id": "INV-001", "text": "AWS Cloud Services March 2025 - Rs 1,24,500 Infrastructure"},
    {"id": "INV-002", "text": "Naukri.com Job Postings Q1 2025 - Rs 45,000 Recruitment"},
    {"id": "INV-003", "text": "Taj Hotels Conference Room Strategy Offsite - Rs 2,85,000"},
    {"id": "INV-004", "text": "Adobe Creative Suite Annual License - Rs 78,000 Design Team"},
    {"id": "INV-005", "text": "EY LLP Consulting Risk Assessment Fees - Rs 8,50,000"},
]

# Build one batch request per invoice
batch_requests = []
for inv in invoices:
    batch_requests.append({
        "custom_id": inv["id"],
        "params": {
            "model": "claude-haiku-4-5",
            "max_tokens": 120,
            "system": (
                "Classify supplier invoices for an Indian enterprise. "
                "Reply in JSON only with keys: "
                "department (IT/HR/Admin/Marketing/Legal/Finance/Operations), "
                "amount_inr (integer), "
                "needs_approval (true if over Rs 200000), "
                "summary (under 8 words)"
            ),
            "messages": [{"role": "user", "content": inv["text"]}]
        }
    })

# Submit entire batch in ONE API call
batch = client.beta.messages.batches.create(requests=batch_requests)
print("Submitted", len(batch_requests), "invoices")
print("Batch ID:", batch.id)
print("Status: ", batch.processing_status)

Polling for batch completion then retrieving and displaying results,

# Poll until processing complete (production: scheduled job checks every 5 min)
def wait_for_batch(batch_id: str) -> None:
    while True:
        b = client.beta.messages.batches.retrieve(batch_id)
        counts = b.request_counts
        print("Status:", b.processing_status,
              "| Done:", counts.succeeded,
              "| Remaining:", counts.processing)
        if b.processing_status == "ended":
            break
        time.sleep(10)

wait_for_batch(batch.id)

# Stream results back - matched by custom_id
print("\n=== INVOICE CLASSIFICATION RESULTS ===")
for result in client.beta.messages.batches.results(batch.id):
    inv_id = result.custom_id
    if result.result.type == "succeeded":
        try:
            data = json.loads(result.result.message.content[0].text)
            flag = " *** APPROVAL REQUIRED ***" if data.get("needs_approval") else ""
            print(inv_id, "->",
                  data.get("department", "?"), "|",
                  "Rs", str(data.get("amount_inr", "?")), "|",
                  data.get("summary", "?"), flag)
        except json.JSONDecodeError:
            print(inv_id, "-> parse error")
    else:
        print(inv_id, "-> failed:", result.result.type)

It gives the following output,

Submitted 5 invoices
Batch ID: msgbatch_01XfGh7kLp2NqR5sWz9yTm3D
Status:  in_progress

Status: in_progress | Done: 3 | Remaining: 2
Status: ended       | Done: 5 | Remaining: 0

=== INVOICE CLASSIFICATION RESULTS ===
INV-001 -> IT         | Rs 124500  | AWS cloud infrastructure costs
INV-002 -> HR         | Rs 45000   | Job posting recruitment platform
INV-003 -> Admin      | Rs 285000  | Hotel conference offsite room  *** APPROVAL REQUIRED ***
INV-004 -> Marketing  | Rs 78000   | Adobe design software license
INV-005 -> Finance    | Rs 850000  | EY consulting risk assessment  *** APPROVAL REQUIRED ***

# 5,000 invoices in one batch = same code, 50% cheaper than real-time calls
# INV-003 and INV-005 flagged for manager approval automatically
# Results available in your DB by morning for the AP team to action at 9am

Batch API workflow for production: each night at 11pm, your ETL job fetches all unprocessed records from the database, builds the batch request list, submits to Claude, and stores the batch_id. At 6am, a second job retrieves results, writes them to the database, and sends a summary email to the relevant teams. The AP team arrives at 9am to a fully classified and flagged invoice queue with zero manual data entry. At 5,000 invoices per month using claude-haiku-4-5, the monthly AI processing cost is under Rs 500 - cheaper than one hour of manual data entry work.

Send your comments, suggestions or queries regarding this site to [email protected].