Anthropic: Claude Opus 5 (batch)

Review this 80-page legal contract for potential liability issues. Flag any clauses that could expose us to uncapped indemnification or IP ownership disputes.

In this illustrative scenario, Claude Opus 5 would process the entire contract within its 1M token window and return a structured analysis identifying three high-risk clauses: Section 12.4's unlimited indemnification language lacking temporal bounds, Section 7.2's ambiguous work-for-hire provisions that could trigger joint ownership claims, and Section 19.1's broad definition of 'Confidential Information' that may inadvertently include your pre-existing IP. The model would cite specific clause numbers, quote relevant excerpts, and explain the legal reasoning behind each flag—demonstrating both document comprehension and domain-aware risk assessment.

The 1M token context window enables full-document analysis without chunking—critical for catching cross-references and contradictory clauses that appear pages apart. At $2.50/Mtok input, batch processing large legal corpora becomes economically viable. However, the batch API introduces latency (typically 12-24 hours), making this unsuitable for real-time contract review workflows.

I'm debugging a memory leak in our React app. Here's the component tree, Redux state snapshots, and Chrome DevTools heap profiles. What's causing the leak?

In this example, Claude Opus 5 would analyze the multi-file codebase submission—component definitions, state logs, and profiler data—to identify that the leak stems from event listeners registered in a useEffect hook without proper cleanup. The model would trace how the listeners accumulate on each re-render, reference the specific line numbers in the submitted files, and provide a corrected code snippet with a cleanup return function. It would also note that the Redux selector in the parent component is creating new object references on every state change, triggering unnecessary child re-renders that compound the listener accumulation.

File upload support and image input (for screenshots of profiler graphs) let engineers submit debugging context in its natural form rather than transcribing everything to text. The batch pricing makes this 10x cheaper than real-time Claude Opus 5, ideal for overnight CI/CD analysis jobs. The trade-off: you can't iterate interactively—submit your full context upfront or wait another cycle.

Translate this 40-page technical whitepaper from English to Japanese, preserving all mathematical notation, figure references, and academic tone. Maintain consistency in terminology across sections.

Claude Opus 5 would process the entire whitepaper as a single context, producing a Japanese translation that maintains terminological consistency—using the same Japanese term for 'gradient descent' in Section 2 and Section 8, preserving LaTeX equation formatting, and adapting figure callouts to Japanese conventions (図1 instead of Figure 1). The model would recognize domain-specific terms like 'backpropagation' and choose the appropriate academic register (誤差逆伝播法) rather than colloquial alternatives. Cross-references between sections would remain accurate because the full document context prevents the fragmentation errors common in chunked translation workflows.

Long-context translation prevents the terminology drift that occurs when translating documents in isolated chunks—the model sees every prior usage. The $12.50/Mtok output cost is significant for a 40-page translation (roughly $15-25 depending on verbosity), but batch processing amortizes this across multiple documents. Real-time Claude Opus 5 would cost 2-3x more for the same job.

When to queue 200+ legal briefs for morning review

A 12-person law firm needs to extract key clauses from 200-300 discovery documents every night. Claude Opus 5 (batch) cuts input cost to $2.50/Mtok—half the real-time rate—and handles the 1M token context window without chunking. Queue the job at 6pm, pull structured JSON by 7am. The trade-off: batch jobs can take up to 24 hours to complete, so this only works if you don't need answers before morning standup. If a partner needs same-day turnaround on a single 80-page contract, switch to the real-time API. For nightly pipelines where cost matters more than latency, batch mode is the call.

How a 5-person startup saves $600/quarter on investor updates

A seed-stage SaaS company generates quarterly board decks by feeding 6 months of Stripe data, support tickets, and product analytics into a single prompt. The 1M token window fits the entire dataset without preprocessing. At $2.50 input and $12.50 output per Mtok, a 400k-token input and 50k-token output costs $1.63 per report—versus $3.25 on real-time Opus. Run it Friday afternoon, review Monday morning. The boundary: if you're generating more than one report per week, the batch discount compounds fast. If you need interactive back-and-forth to refine the narrative, real-time is still faster. For scheduled, high-context summarization jobs, batch pricing makes Opus viable at small-team budgets.

When 50% cost savings justify overnight translation cycles

A 20-person e-commerce team translates 80-120 product descriptions per week from English into 8 languages. Each description includes technical specs, marketing copy, and SEO metadata—roughly 2k tokens per product. Claude Opus 5 (batch) handles nuanced tone and file attachments (PDFs with brand guidelines) at half the real-time cost. Queue 100 jobs Sunday night, publish Monday. The threshold: if your content calendar allows 12-24 hour turnaround, batch mode saves $200-300/month at this volume. If you're launching a flash sale and need translations in 2 hours, pay the real-time premium. For weekly or monthly publishing cadences, batch is the default.