After OpenAI, Anthropic is the second closed frontier lab most teams put on a shortlist. You already met the Claude family in Vol. 11.5 and the wire format in Vol. 18 Anthropic SDK. This lecture is the ecosystem card: Claude chat SKUs, Claude Code (agentic coding CLI/product), Artifacts and Projects on Claude.ai, and the Messages API—not unpublished constitutional internals.
Vol. 21 still owns the product: a Claude backend can power research assistants, document AI, or coding assistants. Module 22.6 covers Claude Code as an IDE/CLI tool in more depth. Here you decide whether Anthropic is your primary vendor, failover, or long-context specialist.
Learning Objectives
By the end of this lesson, students should be able to:
- Name Anthropic’s public surfaces: Claude SKUs, Claude Code, Artifacts, Projects, Messages API.
- Fill a qualitative Pricing / Strengths / Weaknesses card versus OpenAI and Google.
- Separate Claude.ai workspace features from the developer API.
- Map Claude Code to Vol. 21 coding-assistant HITL and Vol. 22.6 tooling.
- Call the Anthropic SDK at a sketch level (system prompt + messages + usage).
- Avoid inventing constitutional-AI internals or fake dollar prices.
Anthropic is a closed-weight frontier lab. Claude is its assistant family (tiered SKUs such as Opus / Sonnet / Haiku-class names—confirm current IDs). Claude Code is Anthropic’s agentic coding product/CLI that operates on a local repo with human review. Artifacts are Claude.ai interactive canvases (code, docs, UI previews) for the consumer/workspace app. Projects are Claude.ai workspaces that attach persistent instructions and files as project context. The API is the Messages API consumed via the official SDK or HTTP. Flagship weights are not generally public.
Product Surfaces
| Surface | Job | Do not confuse with |
|---|---|---|
| Claude chat SKUs | Long-context analysis, writing, tools, coding in API or claude.ai | A single frozen “Claude” checkpoint |
| Claude Code | Repo-aware agentic edits from terminal / IDE-adjacent flows | Unsupervised production deploy without HITL |
| Artifacts | Claude.ai canvas: render/edit generated docs, code, simple UIs | Your production frontend or a public hosting platform |
| Projects | Persistent project instructions + uploaded files for a team workspace | A full RAG platform with your eval/privacy controls |
| Messages API | Metered developer access; tool_use blocks; streaming | ChatGPT-compatible payload copy-paste |
Pricing / Strengths / Weaknesses
Qualitative only. Re-read Anthropic’s current pricing and your evals.
| Axis | Pricing (qualitative) | Strengths | Weaknesses |
|---|---|---|---|
| Claude text / tools | Token-metered API; flagship (Opus-class) above mid (Sonnet-class) above fast/cheap (Haiku-class). Claude.ai Pro/Team/Enterprise are seat subscriptions, not the same as API tokens. Extended thinking / large context burns more tokens. | Strong long-document and careful-refusal product narrative; excellent coding/analysis reputation on many teams’ internal evals; clear Messages contract. | Closed weights; narrower native image/video/ASR catalog than OpenAI/Google; SKU names churn; not automatically cheapest at bulk RAG volume. |
| Claude Code | Often bundled with Claude paid plans and/or API usage depending on current packaging—confirm docs, do not invent seat prices. | High-agency coding loop on a real repo; fits Vol. 21 HITL diff review. | Still needs tests + human ownership; not a substitute for Vol. 22.6 IDE comparison (Copilot, Cursor, etc.). |
| Artifacts / Projects | Included in Claude.ai plans, not a separate public per-token SKU. | Fast internal prototyping and shared project context without standing up RAG. | Workspace features ≠ your multi-tenant product; file/PII governance is still your job (Vol. 20). |
| Enterprise / cloud resale | Anthropic API plus cloud marketplaces (e.g. Amazon Bedrock, Google Vertex, Microsoft Foundry-class listings—availability changes). Commitments, not sticker math. | Procurement via an existing cloud bill; data-residency stories through the cloud partner. | You still do not own weights; dual-path (direct vs Bedrock) complicates logging maps. |
Vs OpenAI
- Often chosen for long context + writing/coding tone
- Fewer first-party still/video/ASR products
- Different tool_use blocks—abstract in your API
Vs Google
- Less native Workspace/Vertex gravity
- Claude not a Veo/Imagen/NotebookLM suite
- Can still appear on Vertex as a third-party model
Pick Claude when
- Document-heavy or coding-heavy Vol. 21 products
- You want Claude Code as the agentic coding surface
- Safety/refusal posture is a stated procurement criterion
Why teams add Anthropic
- Failover and eval diversity vs GPT-only lock-in
- Claude Code + Projects accelerate internal knowledge work
- Messages API is well documented for tools/streaming
Limits to remember
- Closed: no weight-level audit
- Artifacts are not your production host
- Constitutional AI is a public alignment theme, not a spec you can reimplement from this page
Practical API Sketch
Official Python client. System is top-level, not a message role. Confirm current model IDs.
Related Lectures
| Lecture | Role |
|---|---|
| OpenAI | Previous provider; first closed stack |
| Claude family | Model lineage |
| Anthropic SDK | Messages / tools / streaming contract |
| Coding assistants / Claude Code (22.6) | HITL coding product vs this vendor card |
| Research assistants / Document AI | Long-context Vol. 21 fits |
| Google / DeepMind | Next: Gemini + open Gemma + Vertex |
Claude.ai Projects are a full RAG + compliance platform. They are a workspace feature; you still design retrieval, eval, and privacy. Second: Claude Code may ship unreviewed production. Third: Anthropic Messages payloads are drop-in OpenAI clones. Fourth: “constitutional AI” lets you quote secret training recipes. Fifth: seat subscriptions equal API token cost. Sixth: Anthropic replaces OpenAI for images, Whisper, and Sora.
Knowledge Check
- Short Answer: Name Anthropic’s developer API style. Answer: Messages API (system + messages / content blocks).
- True/False: Flagship Claude weights are generally open for download. Answer: False.
- Multiple Choice: Artifacts are primarily: (a) Claude.ai interactive canvases, (b) a video model, (c) Azure OpenAI. Answer: (a).
- Short Answer: What are Claude Projects for? Answer: Persistent project instructions and files in the Claude.ai workspace.
- True/False: Claude.ai Pro billing is the same meter as the API. Answer: False.
- Multiple Choice: Claude Code should be treated as: (a) HITL agentic coding on a repo, (b) unsupervised prod deploy, (c) Whisper ASR. Answer: (a).
- Short Answer: Which Vol. 18 lecture teaches the Anthropic client? Answer: Anthropic SDK.
- True/False: You should copy OpenAI tool JSON unchanged into Anthropic calls. Answer: False.
- Multiple Choice: A common reason to pick Claude over GPT is: (a) long-document / coding / refusal posture on your eval, (b) free unlimited Sora, (c) open GPT-4 weights. Answer: (a).
- Short Answer: Which provider lecture follows Anthropic in 22.1? Answer: Google / Google DeepMind (Gemini, Gemma, NotebookLM, Veo, Imagen, AI Studio, Vertex AI).
Key Takeaways
- Anthropic = Claude SKUs + Claude Code + Claude.ai Artifacts/Projects + Messages API.
- Qualitative pricing: token tiers vs seat plans; thinking/long context costs more; no fake $ tables.
- Strength: long context, coding, careful refusals. Weakness: closed; thinner first-party image/video/ASR suite.
- Abstract the SDK behind your API (Vol. 18). Claude Code still needs Vol. 21 HITL.
- Next: Google / DeepMind.
Lab: Same Vol. 21 product as the OpenAI lab. Students write a Claude vendor card: SKU tier, whether they use API vs Projects vs Claude Code, data path, and one Messages spike. Explicitly list what Artifacts will not replace (auth, multi-tenant storage, eval).
Whiteboard: OpenAI vs Anthropic two-column: surfaces, pricing posture, lock-in. Arrow Claude Code → 22.6. Leave a third column blank for Google.
Recap: Anthropic is the second closed frontier card—Claude, Claude Code, Artifacts, Projects, API. Next is Google / DeepMind.