The Commercial Wedge
Claude Code was the moment Anthropic stopped being a research lab that sold API access and became a platform company that happened to run a safety lab — and the distinction matters more than either characterization suggests.
The standard account of frontier AI commercialization runs: research lab produces capable model, packages it as API, waits for developers to build on top. This is the OpenAI path between GPT-3 and ChatGPT. It generates revenue and surface area, but the lab occupies a specific position in the stack — below the application layer, competing primarily on model quality and price. The developer builds the experience; the lab sells the inference.
Claude Code breaks this pattern in a way that is not primarily about coding assistants. It is about agentic workflows — the category of AI use cases in which the model doesn't respond to a prompt but executes a task over multiple steps, using tools, with state that persists across a session. In agentic contexts, the model is no longer a component that the developer wraps; it is an agent that the developer sets loose. The inference layer and the application layer have collapsed. Anthropic is no longer beneath the developer's product — it is the developer's product.
As of 2026, independent evaluations conducted over multi-week deployments into actual production codebases indicate that Claude Code consistently outranks its competitors in raw intelligence benchmark scores, total context window comprehension, and unassisted autonomous execution capabilities. For many elite senior developers, the optimal converged tech stack involves a hybrid approach: utilizing Cursor for highly localized, immediate text editing while simultaneously deploying Claude Code in the terminal to autonomously execute heavy, architectural multi-file tasks.
The Model Context Protocol is the structural move that compounds the wedge. MCP is an open protocol that specifies how AI models connect to external data sources and tools — file systems, APIs, databases, development environments. Released by Anthropic in late 2024, it is not Claude-specific; other models can implement it. The strategic move is not exclusivity. It is standards-setting. When Anthropic authors the protocol that governs tool use in agentic systems, the protocol becomes part of the infrastructure layer — and Anthropic's models have a structural advantage in the ecosystem the protocol creates, because they were designed alongside it.
The analogy is HTTP. Netscape didn't invent HTTP — the protocol was already there — but Netscape shaped the early ecosystem around it, and the companies that built on that ecosystem had to relate to Netscape's decisions. MCP is Anthropic's attempt to occupy the equivalent position in the agentic layer of the stack. Whether it succeeds depends on whether other major players adopt it, which is why Anthropic made it open and why the protocol's reference implementations include integrations with tools Anthropic doesn't control.
By open-sourcing MCP, Anthropic is attempting to define the Topology of the Agentic Web. If every SaaS platform adopts MCP, then Anthropic controls the "physics" of how agents interact with data. It establishes Anthropic as the "TCP/IP" of the Agentic Era.
The API-first strategy predates both Claude Code and MCP, and it is the constraint that shapes everything downstream. Anthropic's commercial surface is built around API access as the primary revenue driver. Enterprise and API customers are the business; the consumer product (claude.ai) is the proof point and the brand surface, but it is not the financial engine. This is a deliberate choice, not an oversight, and it has structural consequences.
API-first means Anthropic's commercial fate is tied to developers more than to consumers — developers who are sophisticated enough to evaluate model quality on real tasks, developers who are price-sensitive but who also have very high switching costs once a model is embedded in a production workflow. It means Anthropic competes on capability at the API level before it competes on experience at the consumer level. And it means the commercial moat is deepened not by features in the consumer product but by switching costs in the developer ecosystem — by the number of production codebases that call the Claude API, by the MCP integrations that assume Claude as the orchestrator, by the Claude Code workflows that developers have built their tools around.
Anthropic's commercial surface is built around API access as the primary revenue driver. Enterprise and API customers are the business; the consumer product (claude.ai) is the proof point and the brand surface, but it is not the financial engine. This is a deliberate choice, not an oversight, and it has structural consequences: Anthropic's commercial fate is tied to developers more than to consumers — developers who are sophisticated enough to evaluate model quality on real tasks, developers who are price-sensitive but who also have very high switching costs once a model is embedded in a production workflow.
Boris Cherney, who led the Claude Code product, is the right figure for understanding what the commercial wedge actually is. Claude Code is not a coding assistant that uses Claude. It is an agentic runtime that uses Claude as its intelligence layer and exposes the runtime to developers through a CLI. The product's success comes not from what it can do in a single interaction but from what it can do over a multi-step session with access to the developer's codebase — understanding context, planning across files, executing changes, catching errors, and iterating. The session is the product. The model is the intelligence. The wedge is the workflow that neither component creates alone.
Editorial aside: There is a genuine tension between the commercial wedge argument in this chapter and the doctrine argument in chapter 4. If Claude Code's value comes from being embedded in developer workflows, that creates a switching-cost moat that depends on continued capability improvement — which means the commercial success of the wedge depends on Anthropic continuing to ship at the frontier. The RSP's deployment gates become, in this framing, commercial gates as well: a capability Anthropic can't deploy is a capability advantage Anthropic can't monetize. Whether this alignment of incentives makes the doctrine more or less robust is a genuinely open question, and the corpus argues it both ways.
What the commercial architecture reveals about Anthropic's theory of its own position: the company does not believe it can win on consumer experience against Google or Meta. It believes it can win in agentic developer workflows by being the best model for complex, multi-step tasks in the environments where complex multi-step tasks happen — which are, increasingly, developer environments. The wedge is not broad. It is deep. And depth compounds in platform businesses in ways that breadth does not.
The wedge is sharp enough to generate real revenue — and sharp enough to generate real contradictions, which is where the book's honest questions begin.