The Anthropic Stack on One Page

Anthropic ships into so many surfaces now — model tiers, API primitives, builder tools, end-user apps, enterprise controls, and cloud distribution routes — that nobody who isn't paid to track it can hold the whole shape in their head. This post is one point-in-time map of that surface area, grouped by who actually uses each piece. The point is not the categories; the point is being able to see what exists, all at once, on one page.

Six domains. They are not Anthropic's official taxonomy — they are how the product looks to a buyer trying to decide which layer they need to touch. The map is current as of June 2026; the exact model names and preview labels will keep moving.

How to read it

Each card answers a different question for a different audience. Treat the diagram as six conversations stacked on one page.

Engine — which model, on what. Opus when judgment quality is the actual differentiator and the cost per call doesn't matter. Sonnet as the default workhorse for almost everything else. Haiku when latency or unit economics win. Thinking modes are the reasoning controls. Vision means the model can read images. Context window is how much input the model can hold at once — now up to 1M tokens on the larger current models.

Platform primitives — the building blocks an engineer composes. The core endpoints are Messages, Batch, and Files. Model capabilities cover things like streaming, citations, and structured outputs. Context primitives include prompt caching and memory. Tools let Claude call software around it, with computer use still best treated as a beta surface. Each one removes a class of glue code that teams used to write themselves; the more of these you adopt, the less of your own infrastructure is doing work the platform now does for free.

Developer platforms — the surfaces you'd hand a builder. Claude Code is the agentic coding surface many developers live in day-to-day, across terminal and IDE workflows. MCP is the protocol that makes any tool you wire up reusable across clients (covered in API vs. MCP). Skills, hooks, subagents, and slash commands are the customization layer on top of Claude Code. Managed Agents and the Agent SDK are how you run longer, more autonomous jobs.

Surfaces — where people use Claude directly. claude.ai in the browser, the desktop apps on Mac and Windows, the iOS and Android apps, the Chrome extension, IDE integrations, Projects (for persistent context per workspace), Artifacts (for inline outputs you can keep), Cowork (for operator-style assistance), Claude Design (research preview), and the Microsoft 365 add-ins. This is the column most knowledge workers ever see.

Enterprise & governance — the controls that matter once a workflow leaves experimentation. Free, Pro, and Max for individuals. Team and Enterprise for the org. SSO/SAML and SCIM for identity. Audit logs, usage analytics, custom retention, and zero data retention for the compliance team. Fine-tuning is available, but it goes through Bedrock.

Distribution & trust — the deployment-and-compliance answer. Five main routes to run Claude: Anthropic API direct, Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Keep the compliance language precise: SOC 2, ISO 27001, and ISO 42001 are certifications; HIPAA-ready configuration, BAAs, GDPR, and CCPA belong in the regulatory and privacy bucket. The choice between direct API and a hyperscaler is usually about which paper trail the buyer's procurement team already accepts.

Why the map helps

Most product conversations only happen one card at a time. Engineers compare API primitives. CTOs compare Bedrock vs. Vertex. Compliance teams compare SOC 2 footprints. Operators compare what's inside the desktop app. Each of those conversations makes more sense when you can also see the other five — because the answer in one card often constrains the choices in another. Picking Bedrock changes which fine-tuning is available. Adopting Skills changes how Claude Code fits into your stack. Standardizing on Projects changes how team context is packaged; adopting Memory changes what can persist across conversations.

Print the diagram. Mark the four or five chips you actually use. The empty space is the menu of things that are already built and waiting — most of them you can turn on without writing code.