Asger Risom
While Anthropic executives publicly call for AI regulation, the company just merged its tools into an autonomous, token-guzzling superapp designed to take over corporate workflows.
If you listen to the executives running Silicon Valley’s biggest AI labs right now, you would think they are terrified of their own software. Anthropic CEO Dario Amodei has spent months making the rounds in Washington calling for safety guardrails, while researchers regularly warn the public about catastrophic risks. The messaging coming out of the front office is remarkably consistent: This technology is dangerously powerful, and the world needs to slow down.
And then, with complete disregard for their own public hand-wringing, those exact same executives ship massive software updates explicitly designed to embed artificial intelligence into every single corner of the corporate workforce.
Case in point: Anthropic officially eliminated the manual toggle between its standard Claude Chat and its autonomous Cowork agent. As reported by Fortune, the company merged the two into a single, unified assistant. Instead of trusting you to decide whether a prompt requires a quick conversational answer or a multi-step execution loop, the software will now make that call on its own.
This is not a minor user interface cleanup. Anthropic is openly building an AI “superapp”—a single, inescapable interface meant to replace your entire suite of enterprise tools. Alongside the Cowork merger, the company dropped Claude Docs and Claude Slides directly into the main window. The grand vision is a closed loop where you never actually have a reason to leave their platform. You ask Claude a casual question about your quarterly pipeline, and without touching another piece of software, the agent drafts a research document, invites your team to edit it, and formats the results into a presentation deck.
Copying the OpenAI playbook
Anthropic is executing the exact same playbook OpenAI has been running for the last year. OpenAI spent the spring folding its Codex programming agent into the main ChatGPT app, and subsequently scrapped its standalone Atlas browser project to force those agentic, web-browsing capabilities right back into the core product.
The underlying strategy is pure lock-in. If an AI can research, write, code, and design within one continuous session, clicking over to a competitor’s app feels like an unnecessary chore. For the AI companies, this creates a deeply sticky, monopolistic hold on enterprise customers. Once an entire office department is collaborating on live documents inside the chat window, canceling that corporate subscription becomes a logistical nightmare.
The token-guzzling catch
Handing the steering wheel over to an autonomous AI agent also comes with a massive financial catch for the customer: token consumption. AI labs do not bill enterprise clients a flat monthly rate for heavy lifting; they bill based on how many tokens the hardware processes. And agentic workflows are notoriously thirsty.
When you ask a basic chatbot a question, it generates a single response and stops. When you unleash an autonomous agent, it kicks off an endless, recursive loop of tool calls, reasoning steps, and self-correction. According to a Stanford Digital Economy Lab study cited by Fortune, turning a simple coding task over to an autonomous agent consumes roughly 1,000 times more tokens than a standard back-and-forth chat.
By making agentic tool-calling the default machinery behind its new unified interface, Anthropic is setting up enterprise usage bills to climb dramatically. The company insists that users will retain control and that Claude will default to asking permission before taking big actions, but the underlying commercial incentive for the AI provider is obvious.
This is the great theater of the current tech landscape. Silicon Valley CEOs will happily stand in front of Congress and solemnly demand safety legislation to protect humanity from runaway AI. But the moment they walk off stage, they go right back to shipping all-consuming, token-guzzling superapps designed to run as much of corporate America as possible.
