You use Claude's API — a way for your app to talk to Anthropic's AI, like a waiter between your code and the kitchen. Send a prompt, get a response, pay per token (a word-chunk the AI reads, roughly ¾ of an English word). Standard deal. If you don't like it, swap to GPT in an afternoon.
But running AI agents — programs that use AI to complete multi-step tasks on their own — in production requires infrastructure nobody wants to build. Sandboxing, state management, error recovery, scaling. The gap between "works in a demo" and "works on a Tuesday at 3 AM when everything breaks" kills most agent projects before they ship.
On April 8, Anthropic launched Managed Agents — a cloud service that hosts, runs, and manages your AI agents on Anthropic's servers. The price: $0.08 per session-hour (think of it as renting a tiny server for your agent) plus standard token costs. Notion, Rakuten, Asana, and Sentry signed up as early adopters.
Here's what you get for that eight cents. Each agent runs in an isolated container — a sealed box that can't touch anything outside it — with automatic checkpoints (save points so your agent can resume after a crash), scoped permissions (it can only access what you allow), and support for sessions that run for hours. According to SiliconANGLE, Rakuten deployed agents across five departments — product, sales, marketing, finance, HR — within one week per agent. Sentry paired its debugging tool with a Claude agent that now writes patches and opens pull requests (code change proposals submitted for review). Notion integrated agents that let engineers ship code and knowledge workers generate presentations directly in their workspaces.
The timing tells the real story. One day earlier, on April 7, Anthropic hit $30 billion in annualized revenue — annual revenue projected from current monthly earnings — up from $9B at the end of 2025. That's a 3.3x jump in four months, passing OpenAI's ~$25B for the first time. Enterprise customers spending over $1 million per year doubled to more than 1,000 since February. Eight of the Fortune 10 are Anthropic clients. Claude Code — Anthropic's coding tool — alone pulls in $2.5 billion in annualized revenue after nine months on the market. MCP (Model Context Protocol — a universal plug standard for AI tools, like USB but for data) reached 97 million installs by March 2026.
See the pattern? API gives you the model. Claude Code gives you developer tools. MCP standardizes how tools connect. Managed Agents hosts the entire workload. Each layer makes the next one feel like a natural upgrade, not a new vendor decision. It's AWS circa 2008 — start with compute, then make leaving progressively harder. Anthropic's revenue doesn't triple in four months because the model got three times smarter. It triples because each infrastructure layer compounds the one below it.
The tradeoffs are real, and the new ones are worse than the old ones. Managed Agents introduces dual-layered billing — time plus tokens — and nobody knows what a complex agent workflow actually costs until the invoice arrives. A coding agent that runs for four hours, hits three dead ends, backtracks, and finally ships a fix? You're paying for every minute of wandering plus every token of reasoning. With a plain API call, cost scales with output. With session-hours, cost scales with how long your agent thinks — including when it thinks badly. Budget forecasting for agent workloads just became a guessing game dressed up as a pricing model.
But the billing isn't even the sharpest hook. When your agent runs on Managed Agents, its state — checkpoints, execution history, runtime memory — lives on Anthropic's servers. Not your code. Not your prompts. The actual operational context your agent accumulates while doing its job. A Sentry agent that's spent three months learning your codebase's quirks, building implicit context about which files break together, which PR patterns cause regressions? That runtime knowledge exists inside Anthropic's containers. You can export your code anytime. You cannot export what your agent has become. This isn't the old lock-in of switching APIs. This is vendor dependency at the level of institutional memory.
And the session-hour model creates an incentive Anthropic will never mention in a launch blog. Every minute your agent runs is revenue. Efficient agents that solve problems quickly earn Anthropic less than agents that deliberate, retry, and explore. The platform profits from compute time, not from outcomes. Your incentive — fast, cheap results — runs directly counter to the pricing structure. That tension won't matter at $0.08 per hour. It will matter enormously when Managed Agents handles thousands of concurrent sessions and that rate adjusts to "market conditions."
Anthropic sells four layers of AI infrastructure now. The model — the thing you thought you were buying — is just layer one. The tools, the protocol, the hosting stack above it: that's where the margin lives. And with Managed Agents, for the first time, your operational data — not just your API calls, but your agents' learned behavior and runtime state — sits on their servers. You're not renting compute anymore. You're depositing institutional knowledge into someone else's vault and hoping the withdrawal terms stay favorable.
