Anthropic: Claude Sonnet 4.5 (batch)

Review this Python function for potential race conditions. We're seeing intermittent failures in production when multiple workers access the cache simultaneously.

This example would show Claude Sonnet 4.5 identifying a missing lock around the cache read-modify-write sequence, explaining how two threads could interleave between the existence check and the write operation. The response would include a corrected implementation using threading.Lock() with context manager syntax, plus a note about considering Redis or another external cache for true multi-process safety. The explanation would be methodical, walking through the exact interleaving scenario that causes the bug.

Batch processing suits code review workflows where immediate response isn't critical. The 1M token context window means you can submit entire codebases for analysis. Trade-off: 50% cost savings over real-time Claude, but responses arrive asynchronously—typically within hours, not seconds.

Analyze these 47 customer support transcripts and identify the top 3 friction points causing cancellations. Include verbatim quotes and frequency counts for each issue.

This example would demonstrate Claude processing a large batch of conversational data, extracting patterns across dozens of interactions. The output would present three ranked issues (e.g., 'onboarding confusion', 'pricing transparency', 'missing integrations'), each with 3-4 supporting quotes showing the customer's exact language, plus occurrence counts. The analysis would note sentiment shifts and suggest which issue offers the highest-impact fix based on frequency and emotional intensity in the transcripts.

The massive context window handles bulk document analysis that would require chunking in smaller models. Batch mode makes this economical for periodic reporting—run it weekly on accumulated tickets rather than per-conversation. The file input modality means you can upload CSVs or JSON exports directly.

I'm attaching a 40-page technical specification PDF and a 15-page requirements doc. Write test cases covering all security requirements, formatted as Gherkin scenarios with concrete examples.

This example would show Claude parsing both documents, cross-referencing security requirements against technical implementation details, then generating structured Given-When-Then scenarios. Each test case would include specific parameter values, expected error codes, and edge cases derived from the specs. The output would group scenarios by security domain (authentication, authorization, data encryption) and flag any requirements mentioned in one document but not addressed in the other.

Multi-file input plus extended context means no manual copy-paste or summarization before prompting. Batch pricing ($1.50 input) makes processing lengthy technical documents 50% cheaper than real-time alternatives. Best for overnight test generation or documentation audits where you need thorough coverage, not instant feedback.

When batch mode cuts your document pipeline costs by 80%

A 12-person legal tech startup processes 400-600 contracts nightly, extracting clauses and flagging risk terms. Claude Sonnet 4.5 batch handles the 1M token context window they need for full agreements without chunking, and the batch queue cuts their bill from $3,000/month to under $600. Input at $1.50/Mtok means a 200-page contract costs $0.30 to analyze instead of $1.50 on real-time endpoints. The trade-off: results arrive in 12-24 hours, not seconds. If your workflow can tolerate that latency—nightly runs, weekly reports, async research—batch mode is the call. For anything user-facing or sub-hour SLA, stay on real-time Sonnet.

How 1M token context handles cross-file refactors without RAG

A 5-engineer SaaS team needs to rename a core API across 40 TypeScript files and update every call site. Claude Sonnet 4.5's 1M token window fits their entire 180k-token codebase in one prompt—no vector search, no retrieval errors, no missed references. They paste the full repo, describe the change, and get a diff that touches every file correctly. At $1.50/Mtok input, the whole job costs $0.27. The catch: batch mode adds 12-hour turnaround, so this works for planned refactors, not live pairing. If you're doing overnight CI jobs or weekly architecture reviews, the context size and price make this the obvious pick.

When batch pricing makes long-form synthesis viable at scale

A 20-person market research firm generates 80 client reports per quarter, each synthesizing 50-100 source documents into a 15-page narrative. Claude Sonnet 4.5 batch lets them feed 300k tokens of raw data per report and get structured output at $1.50 input / $7.50 output per Mtok. A typical report costs $1.80 to generate instead of $9 on real-time tiers. The file and image modalities mean they can include PDFs and charts directly without preprocessing. The 24-hour batch window fits their editorial calendar—they queue Monday, review Tuesday. If you're producing dozens of long-form artifacts on a weekly or monthly cadence, batch mode turns a budget blocker into a line item.