Claude Batch API — 50% Cost Reduction for Async Work

How Batch API Cuts Costs by 50%

Anthropic charges $3 per million input tokens and $15 per million output tokens for standard requests. Batch API costs exactly 50% less: $1.50 and $7.50 respectively. The trade-off is processing delay—batch jobs typically complete within 24 hours, sometimes faster.

For agents processing thousands of documents, generating training data, or analyzing logs, the cost savings compound quickly. A $100/month Claude bill drops to $50 using batches for non-urgent tasks.

When to Use Batch API vs Real-Time

Real-time API: user-facing features, chat interfaces, synchronous agent decisions, live streaming.

Batch API: background jobs, bulk data processing, daily reports, agent fine-tuning datasets, compliance audits, historical data enrichment, email summarization at scale.

Smart architectures combine both—use real-time for interactive features and batch for everything else.

Building a Next.js Agent with Batch Processing

Queue async tasks in Supabase, submit them as a batch via the Batch API, poll for completion, and webhook results back to your app. This pattern works for any async agent workflow: data extraction, sentiment analysis, code review, or content generation.

The flow: user triggers action → task stored in database → background job submits batch → database polled for status → results processed and returned.

// Next.js API route: submit batch job
import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

export async function POST(req) {
  const requests = req.body.tasks.map(task => ({
    custom_id: task.id,
    params: {
      model: 'claude-3-5-sonnet-20241022',
      max_tokens: 1024,
      messages: [{ role: 'user', content: task.prompt }]
    }
  }));

  const batch = await client.batches.create({ requests });
  
  // Store batch ID in Supabase
  await supabase.from('batches').insert({
    id: batch.id,
    status: batch.processing_status
  });

  return { batchId: batch.id };
}

Monitoring Batch Status and Results

After submission, poll the batch endpoint or wait for webhook notifications. Once complete, retrieve results and store them back in your database. For agents that depend on batch outputs, queue follow-up tasks only after results arrive.

Supabase real-time subscriptions keep your Next.js frontend in sync without constant polling.

Real-World Cost Example

Processing 10,000 customer support tickets monthly for sentiment analysis and categorization. Each prompt: ~500 input tokens, ~200 output tokens.

Real-time: (10,000 × 500 × $3/1M) + (10,000 × 200 × $15/1M) = $45 + $30 = $75/month.

Batch API: same tokens, 50% discount = $37.50/month. Over a year, you save $450 on one workflow alone.

Open-Source Implementation

The Pantheon repository (github.com/lewisallena17/pantheon) provides a production-ready starter kit for batch processing with Claude, Next.js, and Supabase. It includes batch job management, status polling, result webhooks, and error retry logic.

Use it as a foundation for your agent system—remove the boilerplate and focus on your business logic.

Open-source implementation

Everything in this article runs in pantheon — a production-ready Next.js + Supabase + Claude starter. Clone it, deploy to Vercel, run PM2. The dashboard auto-commits every agent edit and reverts itself if TypeScript breaks.