What Anthropic’s Cowork Actually Tells Us About the AI Agent Race

There is a version of this story that is just a product review.

Anthropic has released Cowork, a desktop agent that lives inside your files, works on tasks in the background without you having to hover over it, and notifies you when it needs input or when the job is done. You give it a project brief, and it goes and does the work. That is the headline. That is the press release.

But product announcements rarely arrive in a vacuum, and the one thing the Cowork launch makes unavoidable is the contrast between what Anthropic is building and what its main rival is building. OpenAI is chasing the consumer masses through ChatGPT. Anthropic is methodically constructing an AI infrastructure layer inside the enterprise. These are not the same bet. And the numbers are starting to show which one is paying off.

Revenue tells most of the story. As of mid-2026, Anthropic’s annualized revenue has surpassed $45 billion, a figure that puts it ahead of OpenAI for the first time in absolute terms. The growth trajectory is the more striking part: from roughly $1 billion in ARR in early 2025 to $45 billion fifteen months later. That is not a growth curve. It is a hockey stick. And it did not happen because consumers discovered Claude on their phones.

Anthropic’s revenue is approximately 80% enterprise and developer API. OpenAI’s is roughly 85% consumer ChatGPT subscriptions. These are structurally opposite companies that happen to use the same underlying technology.

The distinction matters more as AI systems move from answering questions to running processes autonomously. When an AI agent operates inside a company’s Slack, CRM, codebase, and document repository simultaneously, the switching cost is not a chat history. It is a full workflow reconstruction. Enterprises do not casually rip that out and replace it with a competitor’s tool, even if the per-token price is slightly lower. Anthropic is building exactly this kind of embedded dependency, and Cowork is the interface through which individual knowledge workers experience it.

What Cowork Actually Does

Wired’s hands-on review described Cowork as an AI agent that actually works, which is a more meaningful endorsement than it sounds. The desktop agent space has produced considerable hype and limited reliable deployment. Agents that plan, browse, write, and execute tasks across multiple applications sound transformative in theory and frustrating in practice, because the gap between a demo and a dependable workflow is wide.

Cowork addresses this differently than most competitors. Rather than giving the agent a chat window and hoping it figures out what you need, you give it a task with a clear deliverable and it manages the process independently. It accesses files on your desktop, reads and writes documents, updates project materials, and comes back with a completed artifact or a specific question. The model is asynchronous: you hand off, you do something else, the agent works, you get a notification.

For knowledge workers, this changes the shape of the workday in small but real ways. Instead of opening Claude, pasting in context, requesting a draft, reviewing it, requesting changes, reviewing again, and iterating, you describe what you need, set the agent working, and return to the output. The iteration loop collapses from a multi-session back-and-forth into a single handoff.

The mobile extension is worth noting separately. Cowork is not only a desktop tool. It connects to mobile apps, can pull information from them, and act on their behalf. For sales teams living inside Salesforce or executives tracking Slack threads across multiple accounts, this extends the agent’s presence beyond the workstation. The practical implication is that a professional can set a task from their phone during a commute and return to a completed output by the time they reach their desk. The always-available assistant model is different from the always-available-on-your-desk model.

The Enterprise Architecture Nobody Else Is Showing

The interesting thing about Cowork is not the feature set. It is the architectural vision underneath it. Anthropic has been deliberately careful about how Cowork ships: it rolled out to Team and Enterprise plans first, with a gradual capability expansion that lets them observe failure modes and correct before mass deployment. This is not the pace of a company trying to capture headlines. It is the pace of a company that has seen what happens when agents are released before the safety scaffolding is ready.

The agent space is producing genuine operational headaches alongside the productivity gains. The Information reported that several large Anthropic enterprise clients discovered they had been overcharged by approximately $1.7 million on their API bills. The mechanism is instructive: when agents run autonomous retry loops inside complex tasks, each failed step generates token-consuming retries that compound quickly. Traditional API usage is bounded by the human at the keyboard. Agent usage is bounded by the complexity of the task, which can be large and unpredictable. Enterprises that deployed agents aggressively discovered that the billing model designed for human-in-the-loop usage did not translate cleanly to autonomous operation.

This is a real problem and also, somewhat paradoxically, evidence that the agents are actually being used. The overbilling happened because the agents were running at scale, executing real workflows, and hitting edge cases that generated excessive retries. You do not accidentally charge $1.7 million on a product nobody is using.

Anthropic’s response to this problem matters. A company that treats the billing confusion as a public relations problem rather than a product design problem will handle it differently than one that treats it as a signal about how autonomous agents need different operational safeguards. The jury is still out on which approach Anthropic takes, but the fact that the problem is becoming visible at all is itself informative. The agent deployment curve is accelerating faster than the infrastructure to manage it.

The Parallel Infrastructure Race

While Cowork was launching in San Francisco, Tencent was releasing Agently Mail in China, giving AI agents their own email addresses with the suffix @agent.qq.com. The timing is coincidental but the direction is not. The global AI industry is simultaneously discovering that autonomous agents need digital identities, communication protocols, and payment rails that do not rely on human intermediaries.

Google’s A2A protocol, now in production at over 150 organizations, defines how agents discover, communicate with, and delegate tasks to each other. The Agent Payments Protocol, co-developed with Mastercard and PayPal, defines how agents authorize and settle transactions between themselves. Stablecoin rails are already handling micro-payments between agents at costs below one cent per transaction, compared to the 50 to 80 cents that credit card rails charge per transaction. A Lowe’s pilot lets an AI agent diagnose a home improvement need, match it to inventory, generate a purchase order, and pay for it via stablecoin settlement, all without a human in the loop.

These are not theoretical constructs. They are shipping infrastructure that is being deployed in production environments right now. The question is not whether agents will communicate and transact with each other. It is how fast this will happen and what guardrails will exist when it does.

The email landscape reflects this acceleration. A joint study by Barracuda, Columbia University, and the University of Chicago found that 51% of malicious emails were AI-generated as of April 2025, crossing the majority threshold for the first time. By December 2025, AI-generated phishing emails had surged 1,265% year over year. The average cost of a phishing-related data breach is $4.88 million per incident, according to IBM’s latest analysis.

The more uncomfortable version of this picture was described in a Cognitive Revolution podcast episode that has been circulating in enterprise AI circles: professionals are increasingly receiving what one guest called pseudo-replies, responses that are grammatically correct and contextually appropriate but sent not by a human but by another person’s AI agent reading and responding to their emails. These professionals have begun manually blocking these reply-robots because their inboxes are filling up with AI correspondence they never asked for.

This is not dystopia science fiction. This is a current working environment for some knowledge workers in 2026. And it is happening before the infrastructure to manage it has fully formed.

The Switching Cost That Changes Everything

Here is the thing about Anthropic’s B2B strategy that the revenue numbers do not directly show.

When a company like KPMG gives Claude access to 276,000 employees across its global audit, tax, and consulting operations, that is not an IT procurement decision. That is an architectural commitment. The integration touches HR systems, client databases, internal knowledge repositories, document management workflows, and compliance logging. Switching away from Claude at that depth of integration means rebuilding all of those connections on a different platform, retraining employees on a different interface, and decommissioning years of accumulated context and institutional memory.

This is why the 70% win rate among new enterprise AI buyers is more significant than it appears. It is not just a preference for Claude’s performance on benchmarks. It is a preference for a platform that has already demonstrated it can be woven into complex organizational workflows without catastrophic failure. The enterprises that have already deployed at scale have already paid the integration cost. The marginal cost of expanding Claude’s role inside those organizations is far lower than the cost of evaluating, procuring, and integrating a competitor.

Anthropic’s Cowork is the latest point in this strategy. It is the consumer-facing surface of an enterprise architecture that has been under construction since the company’s founding. The agent that works in your files without constant prompting is also the agent that gives Anthropic a presence in the daily work habits of individual contributors, which is how technology preferences actually propagate through organizations. People do not adopt tools because their CTO mandates them. They adopt tools because their colleagues are using them and they see the output.

What Comes Next

Cowork will expand. The current feature set is deliberately incomplete, with Anthropic watching how early deployments behave before shipping more capabilities. This pace is slower than what some competitors are moving at, but it reflects a specific theory of the market: the enterprises that survive the agent transition will be the ones that build reliable workflows first, not the ones that deploy the most agents the fastest.

The operational risks are real. The billing model needs redesigning for autonomous workloads. The security and compliance frameworks need to catch up with what agents can actually do inside enterprise environments. The communications infrastructure is forming faster than the governance frameworks that should accompany it.

But the direction is clear. The AI agent that works alongside you as a colleague rather than responding to you as a tool is not a future possibility. It is a present reality, and it is spreading through exactly the route that produces the most durable competitive advantage: deep embedding in how work actually gets done.

OpenAI is building the most famous AI company in the world. Anthropic is quietly building the one that the enterprise cannot afford to rip out.

—
Marcus Webb writes about enterprise AI adoption and the business models shaping the industry’s direction.