Hook: A $100 Billion Valuation on a Mobile Update
Anthropic just unlocked Claude Cowork across all paid plans. Web. Mobile. The same agent that edits files, runs tools, and persists tasks. The announcement came through a crypto news outlet, not a tech blog. That is your first signal. The move is not about technology. It is about narrative.
I have been tracing on-chain lies for seven years. The same pattern repeats here: a product update dressed as a breakthrough, masking the real engineering debt. The code does not lie; only the auditors do. Let me audit this expansion.
Context: The Agent Race
Claude Cowork is Anthropic's answer to the "AI agent" problem. Not a chatbot. A persistent worker that can browse the web, read files, write code, and execute multi-step tasks. It launched in late 2025 as a desktop-only feature, locked behind the $100/month Max plan. Now it is rolling out to Pro ($20/month) and likely Enterprise tiers.
Why now? Because ChatGPT already has mobile, web, and a plugin ecosystem. Google Gemini lives inside Workspace. Anthropic was missing the cross-device surface. This is a catch-up move, not a leap forward.
The source material—a Chinese deep-dive analysis—scores this as a "combined innovation," not an architectural breakthrough. I agree. The core model remains Claude Opus or Sonnet. The novelty is in the orchestration layer: session snapshots, task queues, cross-device state sync.
Core: The Engineering Under the Hood
Let me walk through what this really requires. I have spent years dissecting smart contracts that claim to be "cross-chain." The same problem applies here: state consistency.
When you start a Cowork session on your desktop, the agent initializes a context window. It loads files, sets memory, and begins executing. If you switch to your phone, the agent must pause its current stream, serialize the entire state, and resume on a different device. That is not trivial.
Based on my audit experience, the real challenge is threefold:
- Task Continuity: The agent cannot just stop mid-execution. It must checkpoint its progress—like a database transaction. If the user switches devices, the agent needs to restore the exact state without re-initializing. This requires a robust session store.
- Latency Budget: Mobile devices have limited compute. The agent must offload all inference to the cloud. That means every action—every file read, every tool call—adds network latency. Anthropic must have optimized their inference pipeline to keep response times under 2 seconds on mobile.
- Cost Management: Agent tasks consume 3-5x more tokens than a simple chat. A single Cowork session might run for 10 minutes, executing 50 tool calls. At $0.015 per 1K input tokens, that is roughly $0.75 per session for the company. Multiply that by millions of users, and the monthly cloud bill could hit eight figures.
I do not guess; I verify. The source analysis notes that Anthropic likely has a long-term contract with AWS for GPU capacity. That is a safe bet. But the real question is: did they build a cost-control mechanism that caps the number of sessions per user? Or are they subsidizing usage to grow the user base? The silence on this is deafening.
Silence is the loudest admission of guilt.
Contrarian: What the Bulls Got Right
The bulls will say: this expands the addressable market. Knowledge workers who use multiple devices will now integrate Claude into their daily workflow. Enterprise procurement teams will see cross-platform support as a checkbox item. This could boost monthly active users by 30-40% within six months.
They are not wrong. The move does increase stickiness. Once you train your agent to manage your files, emails, and calendar, switching to a competitor is painful. Anthropic is building a moat through workflow capture.
But here is the contrarian angle: the same workflow capture creates a massive security surface. Mobile devices are stolen, lost, and compromised. If an agent has access to your Google Drive, Slack, and GitHub, a single breach could leak enterprise secrets. The source analysis flags this as a top risk. I agree.
Anthropic's brand is built on safety. But safety in a controlled desktop environment is different from safety on a public WiFi network with a jailbroken phone. The company has not published a mobile security whitepaper. No independent audit of the agent's permissions model. No details on end-to-end encryption.
Every transaction leaves a scar on the ledger. Every mobile session leaves a scar on your data.
Takeaway: The Real Metric to Watch
Forget the hype about "cross-device agents." The metric that matters is cost per task. If Anthropic can keep inference costs below $0.10 per session without throttling, they have a winner. If they cannot, the expansion will quietly become a premium-only feature again.
Watch for a security audit. Watch for a pricing change. Watch for a competitor—either OpenAI or a new entrant—to launch a cheaper, more transparent agent.
Promises are encrypted; data is decrypted. The code does not lie. Only the auditors do. I am watching.