Anthropic: Claude Sonnet 5 (batch)

Review this Python data pipeline for potential race conditions. We're seeing intermittent failures in production when multiple workers process the same S3 bucket.

The model would identify three specific race conditions: (1) the file existence check and subsequent read operation aren't atomic, allowing two workers to both pass the check before either acquires the lock; (2) the temporary file cleanup happens outside the try-finally block, risking orphaned files if the process crashes; (3) the Redis lock has no timeout, so a crashed worker never releases it. It would then suggest using S3's conditional PUT with ETag matching, implementing exponential backoff, and adding lock expiry with a heartbeat mechanism.

Batch processing suits this code-review task well — the 1M token context window handles entire codebases, and the lower cost ($1 input vs $3 for real-time Claude) makes sense when you don't need sub-second responses. The trade-off: batch jobs queue for up to 24 hours, so this only works for async workflows like nightly audits or PR reviews, not interactive debugging.

Extract all financial figures, dates, and entity names from these 47 quarterly earnings PDFs. Output as structured JSON with confidence scores for each extraction.

The model would process all 47 PDFs in a single request (file ingestion support), producing a JSON array where each entry contains extracted revenue figures, EBITDA, guidance ranges, filing dates, and mentioned subsidiaries. Each field would include a confidence score. For ambiguous cases — like a footnote referencing 'adjusted EBITDA excluding one-time charges' — it would flag the extraction with a lower confidence score and include the original text snippet for human review.

This showcases the batch API's strength for high-volume document processing: the 1M token window handles dozens of PDFs simultaneously, and the $1/Mtok input pricing makes bulk extraction economically viable. The model's multimodal capability means it reads tables and charts directly from PDFs. However, batch latency means you're building overnight ETL pipelines, not real-time document chat.

Generate 200 diverse customer support scenarios for testing our new chatbot. Include edge cases: angry customers, ambiguous requests, multilingual queries, and requests outside our service scope.

The model would produce 200 realistic scenarios spanning refund disputes in frustrated tones, vague product inquiries ('I need the blue thing'), queries mixing English and Spanish mid-sentence, and out-of-scope requests like technical support for third-party integrations. Each scenario would include the customer message, expected difficulty level, and suggested ideal response category. The variety would cover seasonal issues (holiday shipping), accessibility needs (screen-reader compatibility questions), and policy edge cases (partial refunds on used items).

Batch processing excels at synthetic data generation — you're trading latency for cost and scale. At $5/Mtok output, generating 200 detailed scenarios costs a fraction of real-time API calls, and the job runs unattended overnight. The context window ensures consistency across all 200 examples (the model 'remembers' earlier scenarios to avoid duplication). The limitation: you can't iterate interactively; you submit the job and wait.

When batch pricing makes 200K-token contract reviews economical

A 12-person legal ops team processes 40-60 vendor contracts nightly, each 80K-200K tokens with exhibits. Claude Sonnet 5 batch cuts their per-document cost to $0.10-$0.25 input (vs. $3.00-$7.50 on real-time Sonnet), making automated clause extraction and risk flagging viable at scale. The 24-hour turnaround window fits their workflow: contracts arrive by 6pm, summaries land in their CRM by 8am. The 1M-token context means full contracts stay in one prompt—no chunking, no lost cross-references. If you need answers in under 4 hours, pay for real-time. If your team reviews contracts during business hours only and can queue overnight, batch pricing turns a $600/month experiment into a $60/month production system.

How vision + batch economics work for high-volume UGC screening

A 5-person community platform moderates 8,000 user-uploaded images daily for policy violations. Claude Sonnet 5 batch handles image + text prompts at $1.00/Mtok input, so a 200-token moderation prompt analyzing one image costs roughly $0.0002—under $2 for the full daily queue. The model's vision capability flags nuanced violations (misleading charts, subtle hate symbols) that keyword filters miss, and the 24-hour SLA works because most content goes live after human review the next morning anyway. If you're moderating real-time chat or live-stream thumbnails, you need synchronous inference. If you're running nightly sweeps on a backlog, batch mode makes multimodal moderation affordable enough to run on every upload instead of sampling.

When Sonnet's reasoning depth justifies batch latency for CFO decks

A 3-person finance team at a Series B startup generates 15-page board decks monthly, pulling data from Stripe, QuickBooks, and Salesforce exports. Claude Sonnet 5 batch ingests 60K-token CSVs, writes narrative analysis of burn rate and cohort trends, and formats tables—all for $0.06 input + $0.30 output per deck. The Sonnet reasoning layer catches anomalies (sudden CAC spikes, unusual refund clusters) that template tools miss, and the 1M-token window means they can feed 12 months of transaction history in one prompt for YoY comparisons. The 24-hour batch window works because decks are due 48 hours before board meetings. If you're building real-time dashboards, use a faster model. If you're generating monthly artifacts where quality beats speed, batch Sonnet turns a $15 task into a $0.36 task.