Anthropic: Claude Opus 4.6 (batch)

Review this 50-page technical specification document for our API redesign. Flag any inconsistencies in authentication flows, identify sections where error handling isn't fully specified, and suggest three architectural improvements.

This example would show Claude Opus 4.6 processing the entire document in one pass, producing a structured analysis with specific page references. The response would include: a table mapping six authentication flow inconsistencies across sections 3.2, 7.1, and 12.4; detailed callouts for eleven endpoints missing error code specifications; and three architectural recommendations (connection pooling strategy, rate limit design, webhook retry logic) each with implementation trade-offs. The analysis would maintain context across all 50 pages, cross-referencing related sections without losing thread.

The 1M token context window makes this model viable for whole-document analysis that smaller models can't handle in one shot. Batch processing keeps costs manageable at $2.50/Mtok input for this workload. The output latency trade-off means this isn't suitable for real-time review sessions—best for overnight analysis runs.

Here's a screenshot of an error message from our deployment pipeline. Explain what's failing, why it's happening based on the stack trace, and give me the exact fix with file paths and line numbers.

This example would demonstrate Claude Opus 4.6 parsing the screenshot to extract the stack trace, identifying a Kubernetes ConfigMap mounting issue in the CI/CD workflow. The response would explain that the deployment is failing because the config volume reference in `deploy/k8s/production.yaml` line 47 points to a ConfigMap that doesn't exist in the target namespace. The fix would include the exact kubectl command to create the missing ConfigMap and the corrected YAML snippet with proper namespace qualification.

Image understanding combined with deep technical reasoning makes this model effective for visual debugging tasks. The batch API's asynchronous nature means you'd queue these during incident response rather than get instant answers—acceptable when you're documenting issues for later resolution, less so during active outages.

I'm attaching our Q3 customer interview transcripts (28 files, ~400 pages total). Extract the top 10 feature requests by frequency, group them into themes, and for each theme identify which customer segments are asking and what their stated use case is.

This example would show Claude Opus 4.6 processing all 28 transcripts to produce a ranked analysis. The output would list ten feature requests with exact mention counts, grouped into four themes (workflow automation, reporting flexibility, mobile access, integration depth). For each theme, the response would map requesting segments (enterprise vs. SMB, technical vs. business users) and quote specific use cases from the transcripts with customer identifiers. The analysis would note patterns like enterprise customers requesting SSO while SMB customers want simpler onboarding.

Multi-file synthesis across hundreds of pages is where the massive context window justifies the higher output cost ($12.50/Mtok). Batch processing makes this economical for periodic research analysis. The model can hold all transcripts in context simultaneously rather than summarizing incrementally, which preserves nuance that chunked approaches lose.

When batch mode cuts your document pipeline costs by 50%

A 12-person legal ops team processes 200+ vendor contracts nightly—extracting clauses, flagging risks, generating summaries for morning review. Claude Opus 4.6 in batch mode is the right call here. You're paying $2.50/Mtok input instead of the real-time $15, and the 1M token context window means each contract (typically 40-80k tokens with exhibits) fits in a single call without chunking logic. The trade-off: batch jobs return results in 12-24 hours, not seconds. If your workflow is asynchronous—documents uploaded by 6pm, results needed by 8am—that latency is invisible and you're saving $12.50 per million tokens of output compared to real-time Opus. Below 50 documents/day the savings don't justify the batch integration work; above that threshold, you're looking at $600-1200/month in cost reduction for the same model quality.

Why 1M tokens matters for cross-repo refactoring tasks

A 5-engineer startup needs to migrate a Python monorepo from Flask to FastAPI—touching 80+ files, 300k tokens of code, plus 150k tokens of dependency docs and migration guides. Claude Opus 4.6 handles this in one context window where smaller models force you into fragile RAG pipelines or multi-turn orchestration. You load the entire codebase, the FastAPI docs, and your internal style guide, then ask for a refactor plan with file-by-file diffs. The batch pricing ($2.50 input) makes this economical even for daily runs during the migration sprint. The constraint: if you need sub-5-second responses for interactive debugging, batch mode's 12-hour turnaround breaks the workflow. For overnight analysis jobs that inform the next day's sprint, this is the highest-context, lowest-cost option in the Opus family.

When image+file support closes the spreadsheet-to-narrative gap

A 20-person finance team at a Series B SaaS company generates board decks quarterly—combining Excel exports, chart screenshots, and 60-page transaction logs into a narrative summary with risk callouts. Claude Opus 4.6's multimodal input (text, image, file) means you can feed raw spreadsheets, PNG charts, and PDF statements into a single batch job without pre-processing them into text. The 1M token window holds an entire quarter's data, and the $12.50/Mtok output pricing is manageable when you're generating 15-20k tokens of prose four times a year. The boundary: if you're running this monthly or need real-time iteration with the CFO, the batch latency (12-24 hours) and lack of streaming output become friction points. For quarterly cadence where a human reviews and edits the next day, batch Opus delivers Opus-tier reasoning at half the cost of real-time.