Anthropic: Claude Sonnet 4.6

Review this pull request for security issues. The code adds user authentication but I'm worried about the session handling and password storage approach.

This implementation would illustrate Claude Sonnet 4.6 identifying three critical vulnerabilities: passwords stored with MD5 hashing instead of bcrypt, session tokens generated with Math.random() rather than cryptographically secure methods, and missing CSRF protection on the login endpoint. The response would provide specific line references, explain the attack vectors each flaw enables, and offer concrete remediation code using industry-standard libraries. The explanation balances accessibility for junior developers with technical precision for security-conscious reviewers.

Security review showcases the model's ability to parse code context across multiple files and apply domain expertise. The 1M token context window means entire codebases can be analyzed together rather than file-by-file. However, without public benchmark data, we can't compare its vulnerability detection rate against specialized security models.

I'm attaching our Q3 financial statements, last year's annual report, and the new product roadmap deck. Write an investor update email explaining how our burn rate aligns with the roadmap timeline.

This example would show Claude Sonnet 4.6 synthesizing numerical data from the financial statements, strategic context from the annual report, and milestone dates from the roadmap into a cohesive 4-paragraph email. The output would include specific figures (current runway in months, projected revenue milestones) while maintaining appropriate tone for investor communications. The model would flag any timeline mismatches between cash reserves and planned product launches, demonstrating cross-document reasoning.

Multi-document synthesis leverages both the extended context window and multimodal file processing. At $15/Mtok output pricing, this task costs roughly $0.06 for a 4000-token response—reasonable for high-stakes communications but expensive for routine drafting. The model's ability to work with uploaded files removes the need for manual copy-paste.

Explain how gradient descent works to a product manager who needs to understand why our ML model's training is taking longer than expected. Use an analogy, then connect it to compute costs.

This response would illustrate Claude Sonnet 4.6 opening with a 'hiking down a foggy mountain' analogy—small steps, checking slope, adjusting direction—then mapping each element to learning rate, loss function, and iteration count. The explanation would transition to practical implications: why larger datasets require more iterations, how batch size affects memory and speed trade-offs, and what 'convergence' means for the training timeline. The model would avoid both oversimplification and unnecessary jargon, calibrating technical depth to the stated audience.

Technical translation tasks highlight the model's audience adaptation and analogical reasoning. The lack of public benchmarks means we can't quantify its performance on explanation quality metrics, but the Sonnet tier historically balances speed and capability well for this use case. The $3 input pricing makes iterative refinement of explanations economical.

When 1M-token context makes legal review actually work at scale

A 12-person legal ops team needs to cross-reference clauses across 40+ vendor agreements before every renewal cycle. Claude Sonnet 4.6's 1M-token window means you load the entire contract portfolio into one session and ask comparative questions without chunking or retrieval hacks. At $3/Mtok input, a full portfolio review costs under $10—cheaper than the engineer-hours you'd burn building a RAG pipeline. The $15/Mtok output rate stings if you're generating full redlines, but for Q&A and clause extraction it's negligible. If your team runs fewer than 100 contract sessions per month, this is the model. Above that volume, you're paying $1,500+/month on output alone and should evaluate a cheaper long-context alternative like Gemini 1.5 Pro.

Why image-plus-text input justifies the output premium for CX teams

A 20-seat SaaS support team gets 300 tickets daily, half with screenshots of broken UI states or error messages. Claude Sonnet 4.6 ingests the image and the user's description in one call, then writes a Zendesk macro or routes to engineering with structured context. The multimodal capability eliminates the "please describe what you see" back-and-forth that kills CSAT scores. At 200 tokens average output per ticket and 300 tickets/day, you're spending roughly $270/month on generation—worth it if each saved exchange is 4 minutes of agent time. The missing benchmark data means you're flying blind on accuracy versus GPT-4o, so run a 2-week A/B test on triage precision before committing your whole queue.

When to skip this model for high-throughput summarization work

A 4-person research team processes 500 academic PDFs per week into 150-word summaries for a lit review database. Claude Sonnet 4.6 handles the file input natively and the 1M-token window means even 80-page papers fit in one shot, but the $15/Mtok output rate makes this a $1,125/month job at 500 summaries/week (assuming 150 tokens each). Compare that to Gemini 1.5 Flash at $0.30/Mtok output: same task costs $22.50/month. Unless you need Anthropic's specific safety filtering or you're already locked into their API for other workflows, the 50x cost delta doesn't pencil out for bulk summarization. Use this model for the 10% of papers that need multimodal figure analysis; route the rest to a cheaper long-context option.