← All days

Day 4

Know the product you’re building on — Claude’s current capabilities, strategic positioning, and the full product surface.

Context

Anthropic's Claude family as of early 2026 has three tiers: Claude Haiku 4.5 (claude-haiku-4-5-20251001 — fastest, cheapest, ideal for high-volume classification and extraction), Claude Sonnet 4.6 (claude-sonnet-4-6 — the production workhorse, best performance-to-cost ratio, the model most products are built on), and Claude Opus 4.6 (claude-opus-4-6 — most capable reasoning, used for complex analysis where quality justifies cost). Model specifications — context windows, output token limits, and pricing — change with updates. Always verify at docs.anthropic.com/en/docs/about-claude/models before making any product decision. Hardcoding specs from a course is how you end up with stale architecture documents.

Claude's capabilities have expanded significantly in 2024-2025 beyond basic chat: Extended thinking (the model generates internal chain-of-thought for harder problems, configurable via a budget_tokens parameter), web search (available via tool use in the API and natively in Claude.ai), Projects (persistent context containers in Claude.ai for teams), computer use (controlling GUIs — covered on Day 25), and Claude Code (an agentic coding CLI — covered on Day 43). Understanding which capabilities are available in which product surface is critical PM knowledge.

A PM at Anthropic needs to understand that Claude.ai (the consumer/team/enterprise product) and the Claude API (the developer product) have different feature sets and release cadences. Features sometimes ship in Claude.ai first (e.g., Projects, web search) and in the API later, or vice versa. The full Anthropic product surface in 2025-2026: Claude.ai (Free, Pro, Team, Enterprise tiers), Claude API (direct access), Claude on AWS Bedrock, Claude on Google Cloud Vertex AI, Claude Code (agentic CLI), and Claude embedded in partner products (Salesforce, etc.). Each channel has different security guarantees, pricing, and feature availability. Knowing the differences is how you recommend the right deployment path for each customer.

Anthropic's mission — "the responsible development and maintenance of advanced AI for the long-term benefit of humanity" — is operational, not decorative. The RSP (Responsible Scaling Policy, covered Day 19) ties commercial deployment to safety evaluations. For PMs, this means Anthropic will sometimes say "not yet" to capabilities that are technically possible but not safe to deploy. That constraint shapes what you can build and when — understand it as a feature of the platform, not a limitation.

Claude's strongest areas relative to competitors: long-context understanding, instruction-following fidelity (especially structured output compliance), coding (explanation, review, generation), nuanced refusal handling (reasoning-based via CAI, not keyword-matching), and safety-by-design (enterprise customers in regulated industries value this). Claude's historical limitations — no web access, no image generation — have partially evolved: web search is now available via tool use in the API, though image generation remains outside Claude's capabilities. Knowing both strengths and limitations honestly is essential for enterprise sales credibility.

Tasks (4)

  1. Build a model selection decision tree (25 min)
    Create a flowchart: given a use case, which Claude model do you choose? Decision factors: expected daily request volume, required latency (p99 ms), context length needed, whether extended thinking is required, budget per 1M tokens. Look up current specs at docs.anthropic.com/en/docs/about-claude/models (don’t use hardcoded numbers). Test on 5 product scenarios. Save as /day-04/model_selection_decision_tree.md.
  2. Research Anthropic’s hiring to infer strategy (25 min)
    Go to Anthropic’s careers page (anthropic.com/careers). What are the top 5 open roles by volume? What does the hiring pattern tell you about where Anthropic is investing? AI safety research, product engineering, enterprise sales, Claude Code? Write a 150-word interpretation. Save as /day-04/anthropic_hiring_analysis.md.
  3. Map Anthropic’s full product surface (25 min)
    Create a table mapping Anthropic’s product surface: Claude.ai (Free/Pro/Team/Enterprise), Claude API, AWS Bedrock, Google Vertex AI, Claude Code, partner integrations. For each: target user, key differentiator, and security/compliance characteristics. When would an enterprise customer choose Bedrock over direct API? Save as /day-04/product_surface_map.md.
  4. Draft honest enterprise positioning talking points (25 min)
    Write 3 reasons an enterprise customer should choose Claude over GPT-4o, AND 2 scenarios where GPT-4o or Gemini might be a better fit. Be specific, honest, and focused on use-case-level tradeoffs. Include the open-source objection from Day 1. Save as /day-04/enterprise_positioning_talking_points.md.

Interview question

Why would you choose to build on Claude rather than GPT-4o?

The answer depends entirely on the use case, but here are the scenarios where Claude wins and where it doesn’t.

Claude wins on: First, long-context applications. Claude Sonnet 4.6’s 200K context handles full contracts, codebases, and research papers without chunking — a genuine architectural simplification over models with smaller windows. Second, instruction-following fidelity. For products requiring strict structured output (JSON, specific formats, compliance-sensitive document generation), Claude consistently follows complex instructions with fewer deviations. Third, safety-by-design for regulated industries. Constitutional AI training means Claude reasons about edge cases rather than pattern-matching, which matters for products touching legal, medical, or financial content. Fourth, enterprise deployment flexibility: Claude is available on AWS Bedrock and Google Vertex AI, matching where the customer already has cloud agreements.

GPT-4o wins on: Largest developer ecosystem (more third-party tools built for OpenAI first), Microsoft integration depth for Azure-native enterprises, and the o-series reasoning models for tasks requiring multi-step mathematical or scientific reasoning. Gemini 2.5 Pro wins on context window (1M+ tokens) for extremely long document use cases.

My principle: always benchmark both on your specific use case before committing. Public benchmarks don’t predict domain-specific performance reliably. And know the open-source alternative: "Why not self-host Llama 4?" has a specific answer (SOC 2, HIPAA, SLA, support, Constitutional AI) that you should be able to recite.

PM angle

At Anthropic, you will be the person who explains your product most honestly — including its limitations. PMs who articulate both where Claude wins and where it doesn’t are more trusted by customers and engineers than those who just sell. And knowing the full product surface (Claude.ai vs API vs Bedrock vs Vertex vs Claude Code) lets you recommend the right deployment for each customer rather than a one-size-fits-all answer.

Resources