Anthropic: Claude Opus 5

Review this TypeScript function for potential race conditions. It fetches user data, updates a cache, then triggers a webhook. The cache write sometimes fails silently in production.

The function has a classic time-of-check-to-time-of-use race. Between reading `cache.has(userId)` and writing `cache.set(userId, data)`, another request could insert the same key. More critically, your webhook fires before confirming the cache write succeeded—if `cache.set()` throws asynchronously or returns a rejected promise you're not awaiting, the webhook sends stale state. Wrap the cache operation in try-catch, await the promise, and only trigger the webhook after successful persistence. Also consider using a compare-and-set primitive if your cache supports it.

Demonstrates deep reasoning about concurrency and error propagation across a multi-step async flow. The 1M token context window means you could paste an entire codebase here for cross-file analysis, though this example shows the model doesn't need that scale to catch subtle bugs. At $25/Mtok output, lengthy explanations add up—concise prompts work better.

I'm pitching a SaaS dashboard redesign to a CFO who hates change. Write three bullet points that frame the redesign as risk reduction, not feature work. Our current UI causes 40% of support tickets.

• **Cutting support load by 40% returns 1.2 FTE to revenue work** — the redesign pays for itself in 8 weeks by eliminating repetitive ticket resolution tied to the current UI's ambiguous CTAs. • **Reduced user error = reduced churn risk** — when 40% of tickets stem from UI confusion, we're training customers to expect friction. Competitors with clearer interfaces become attractive during renewal. • **Design debt compounds** — delaying this work means the next feature launch inherits the same confusion patterns, multiplying future support costs and lengthening your team's iteration cycle.

Shows strong audience modeling—the framing avoids design jargon and leads with financial impact. The model inferred the CFO's priorities (cost, risk, ROI) from a single sentence of context. Multimodal input means you could attach the current UI screenshot and competitor examples for even tighter messaging, though text-only prompts already produce sharp business writing.

Explain quantum entanglement to a bright 14-year-old who just finished algebra but hasn't seen probability theory yet. Use an analogy that doesn't mislead about the physics.

Imagine you and a friend each get a mystery box. You open yours in London, she opens hers in Tokyo, and somehow you always get opposite results—if yours shows red, hers shows blue, every single time, instantly. Entanglement is like that: two particles share a connection so that measuring one immediately determines the other's state, no matter the distance. Here's the non-spooky part: it's not that one box sends a signal to the other. Instead, the pair was created in a way that locked their fates together from the start. You can't use this to send messages faster than light, because you can't control what your box shows—you only see the correlation after comparing notes.

Balances accessibility with precision, avoiding the common 'telepathy' misconception while staying within the constraints of the audience's math background. The model adapts tone and complexity on the fly. However, for highly specialized pedagogy (e.g., curriculum design), a domain-specific fine-tune might better match institutional style guides. General-purpose LLMs excel at one-off explanations, less so at systematic lesson sequencing.

When Claude Opus 5 handles 200-page merger reviews in one pass

A 12-person legal ops team needs to cross-reference clauses across acquisition documents totaling 180-220 pages per deal. Claude Opus 5's 1M token context window fits the entire document set in a single prompt, letting you ask "flag every indemnification clause that conflicts with Section 7.3 of the master agreement" without chunking or retrieval overhead. At $5 input per million tokens, a 200-page review costs roughly $2.50 in input fees—cheaper than the engineering time to build a RAG pipeline. The $25/Mtok output rate stings if you're generating 50k-token summaries, but for extraction and flagging tasks where output stays under 5k tokens, total cost per deal runs $4-6. If your team closes 8+ deals per month and speed matters more than penny-pinching on output, this is the call.

Why Opus 5 is overkill for tier-1 support but right for executive escalations

A 40-person SaaS company routes angry enterprise customers to a dedicated escalation queue where one wrong answer costs the account. Claude Opus 5's text and image modalities let the agent parse screenshots, error logs, and multi-threaded email chains in the same prompt, then draft a response that references all context without hallucinating details. The model's cost—roughly $0.15 per complex escalation at typical prompt sizes—is noise compared to a $120k annual contract at risk. For tier-1 volume (500+ tickets/day), the math breaks: you'd spend $75/day on a workload where a $15/day model performs identically. The threshold is roughly 50 high-value interactions per month. Below that, route to a cheaper model; above it, Opus 5's accuracy and context handling justify the premium.

When Opus 5 turns 6-hour architecture meetings into same-day RFCs

A 9-engineer platform team records 90-minute architecture discussions, uploads the transcript plus 12 pages of prior design docs, and asks Claude Opus 5 to draft an RFC capturing decisions, trade-offs, and open questions. The 1M token window holds the transcript (18k tokens), historical context (40k tokens), and related Slack threads (15k tokens) without summarization loss. Output cost is the pinch point: a 4k-word RFC costs roughly $0.50 in output tokens, and if you're iterating 3-4 times per doc, you're at $2 per finalized RFC. For a team shipping one major design per sprint, that's $50/year—trivial. For a team writing 50 RFCs per quarter, it's $400/quarter, and you should benchmark against a $15/Mtok output model to see if quality delta justifies 67% higher cost. If your RFCs go to the board or determine 6-month roadmaps, pay for Opus 5.