Anthropic merges Claude and Cowork into a single app: the race for the AI 'superapp' hides a bill that could balloon

By SiliconANGLE · September 16, 2026.

Anthropic has merged its Claude chatbot with Claude Cowork, the agentic tool capable of carrying out multi-step tasks on the user's behalf, with immediate effect. As the company explained on its blog, there will no longer be any need to pick a tab: Claude will automatically decide whether a query should be handled by the regular chat or by Cowork's agentic machinery, and context will travel between the two modes on its own.

Alongside the merger come two new beta features built directly into the conversation: Claude Docs, for drafting and editing text documents with real-time collaboration between several people and export to Word or Google Docs, and Claude Slides, for creating, editing and presenting slides from within the chat, with downloads in PowerPoint or PDF. They are joined by Claude Design, until now a standalone product, which likewise moves inside the chat. The rollout starts with paying subscribers —Claude Pro and Claude Max, on web, desktop and mobile— and will reach Team plans and free users in the coming weeks. One detail that is no accident: Claude Code, Anthropic's coding agent, is left out of this move and remains a separate product.

Anthropic's stated goal is to end the friction of having to know in advance which tab or which product to use for each task. It is a real nuisance: today a user who wants to draft a report, look up a fact and automate a repetitive task may have to go through three different tools, each with its own context window. The promise is that writing will be enough and the system will sort out the rest underneath.

This is not just a usability improvement. Consolidating chat, agent, documents and presentations into a single surface builds what the industry already calls a "superapp": a single point of entry that absorbs more and more of the user's working day and, along the way, makes it far harder to migrate to a competitor. The more tasks Claude handles without leaving its window, the more it resembles an operating system for work and the less an app among others. For companies, moreover, centralizing AI use in a single tool means more control over what employees do with these systems, which makes Claude easier to sell in the corporate segment. And for Anthropic, the more features depend on a single subscription, the easier it becomes to bundle and raise prices without users feeling they are paying for separate products.

OpenAI is following the same path: it is believed to be working on integrating ChatGPT, its Codex coding agent and possibly its Atlas browser as well into a single product, with the same argument about reducing tool fragmentation. It is no coincidence that the two labs are fighting this second battle, the one over distribution, in parallel. The most capable model is no longer enough to win: you have to be the place where people choose to work every day. It is the same dynamic we already saw in Google's and Microsoft's approach to agent integration: competitive advantage shifts from the lab that trains the model to whoever controls the interface and the "plumbing" that connects the user to that model.

There is, however, some fine print worth not overlooking. Agentic systems like Cowork reason in several steps and make tool calls on the user's behalf, which consumes far more tokens —the units of text the provider bills for— than a simple chat exchange. If Claude's automatic router decides, by mistake or by conservative design, to treat a simple question as an agentic task, the user may end up paying considerably more without having changed how they use the tool or having consciously decided to. It is a risk that deserves watching, especially as the feature extends to more plans.

Our reading is that this move confirms something we have been pointing out: competition between the big AI labs is no longer fought only on benchmarks, but over who gets users to stop thinking about which tool to use. It is a natural phase of product maturity —moving from a collection of loose utilities to a unified layer of work— and, over the long term we defend at Zendoric, it is consistent with the idea that AI should disappear as friction and become infrastructure: the less the user has to think about the tool, the more attention they can devote to what really matters to them. But in the short term, the fine print of per-token billing is exactly the kind of detail that separates a superapp that empowers from one that overbills by default. It is worth Anthropic being as transparent about the cost of automatic routing as it has been about its convenience.