Anthropic doesn't ship one Claude anymore — it ships a family of surfaces, each built for a different job. Ask "which Claude should we use" and you'll get five different correct answers depending on whether you're chatting, building, or running a pipeline unattended. For a team evaluating where an AI-assisted content operation should actually live, the surfaces matter more than the model name, so here's what each one is for.
Claude.ai — the surface for a person typing
Claude.ai is the consumer chat product: Free, Pro, Max, and the seat-based Team and Enterprise plans. It's subscription-priced and human-driven — someone opens a browser tab, writes a prompt, reads the answer. It's excellent for research, outlining, and one-off drafting help, and it's where most people first meet Claude. It is not built to run unattended: usage is capped by the plan rather than metered, and there's no scheduling or pipeline layer under it.
The Claude Developer Platform (API) — the layer a pipeline runs on
The API is what an actual content engine is built on: pay-per-token access to the model family, callable from code, with no person required to click "send." As of mid-2026, Claude Sonnet 5 runs $2 per million input tokens and $10 per million output tokens — Anthropic's most agentic Sonnet to date, capable of planning and using tools autonomously. Claude Opus 5, released a month later, targets the harder judgment calls at a lower price than the previous Opus generation, with a million-token context window for cases that need it. This is the tier that makes a topic-to-published-post pipeline arithmetically sane: token cost per article, not seat cost per person.
Claude Code and Claude Code on the web — where the automation executes
Claude Code is Anthropic's own coding agent, run from a terminal or an IDE. Claude Code on the web extends that into the browser: point it at a GitHub repository, choose a network-access level for the sandbox (locked down, an allow-list of domains, or open), and give it a prompt. The session runs in an isolated cloud container, can work for extended periods, and opens a pull request when it's done. For a publishing pipeline, this is the environment that builds and maintains the pipeline software itself — the scripts that call the API, not the API calls a marketer makes by hand.
Claude Agent SDK — build the orchestrator yourself
For teams that want more control than a hosted product gives them, the Claude Agent SDK is the same agentic framework Claude Code is built on, exposed for custom use: conversation management, tool and MCP integration, and the harness logic for a long-running agent. It's the right layer when off-the-shelf scheduling and tool access don't fit the workflow — a custom topic-selection loop pulling from a client's own analytics, for instance — rather than a general chat interface.
Claude Cowork and Claude Tag — for the people who don't write code
Claude Cowork is a desktop agent aimed at non-developers working with files and documents; its scheduled-tasks feature lets someone write a prompt once, set a cadence, and let it run without touching a terminal. Claude Tag is the current Slack integration — a shared teammate that joins a channel so anyone can tag @Claude with a request, rather than the earlier personal-assistant version tied to one person's account. Both matter for the same reason: they put a version of the pipeline in front of a marketer or an account manager who will never open the API docs.
Amazon Bedrock and Google Vertex AI — the same models, a different billing and compliance stack
Bedrock and Vertex AI don't add capability; they add routing. The models are the same ones available directly from Anthropic, but authentication, audit logging, and billing run through AWS IAM/CloudWatch or GCP IAM/Cloud Audit Logs instead of a separate Anthropic account. This matters when a client's infosec team has already drawn a hard line around one cloud, and it's the right reason to choose it — not a reason to expect different output from the model underneath.
So which environment does a content operation actually need?
In practice it's rarely one answer. The pipeline itself — briefs, drafts, quality gates, publishing — runs on the API, because that's the only tier priced and built for unattended volume. Claude Code and Claude Code on the web are where that pipeline gets built and maintained, not where the content runs day to day. Cowork and Claude Tag are the right home for an individual marketer's own recurring task, not the whole engine. And Bedrock or Vertex only enter the picture when a client's existing cloud commitments require it. Picking the wrong layer is how teams end up either overpaying for seats to do a machine's job, or gluing chat exports together instead of building something that actually runs.