Anthropic doubles Claude usage limits and begins agent ‘dreaming’ tests

Anthropic is making a pretty direct bet on what comes next for AI — less chatbot, more worker. The immediate news is simple: on May 6, it doubled Claude Code’s five-hour usage limits for paid users and raised API limits for Opus, because it says it suddenly has a lot more compute to work with. But the more interesting part is what Anthropic seems to be building on top of that headroom — agents that can keep working longer, remember more, and even improve themselves between runs. ### What actually changed in Claude? Anthropic’s May 6 announcement was mostly about capacity. Claude Code’s five-hour rate limits were doubled for Pro, Max, Team, and seat-based Enterprise plans. Peak-hour reductions for Pro and Max were removed. And API rate limits for Claude Opus models were increased, which matters for developers running heavier automated workflows instead of one-off chats. (anthropic.com) ### Why could Anthropic suddenly do that? Because it signed a new compute partnership with SpaceX. Anthropic said it will use all of the capacity at SpaceX’s Colossus 1 data center, giving it access to more than 300 megawatts of new capacity and over 220,000 Nvidia GPUs within the month. Basically, this is the infrastructure story behind the product story — more chips means fewer artificial bottlenecks on coding agents and API-heavy enterprise use. (anthropic.com) ### Why is Claude Code the center of this? Because coding agents are where usage limits hurt the most. A normal chatbot session is bursty. A coding agent can run tools, inspect files, retry, test, and keep going for hours. Anthropic has been leaning hard into that market for a while, with Claude Code, Computer Use, and a steady push toward agentic workflows. If you want developers and companies to trust an agent with real work, the agent cannot keep hitting the wall halfway through the job. (anthropic.com) ### So what is this “dreaming” thing? It looks like Anthropic’s label for self-improvement loops in managed agents. The basic idea is that an agent does a task, reviews what went wrong, stores the useful lesson, and comes back better next time. That is not the same as retraining the base model. It is more like giving the agent a way to write better playbooks for itself. Anthropic has also been talking publicly about self-improving agents with partners like Warp, which makes the direction pretty clear even if the feature is still early. (anthropic.com) ### Why does that matter more than a bigger context window? Because enterprise buyers mostly do not need a smarter demo. They need a system that gets less flaky over time. The big unsolved agent problem has been reliability — agents forget instructions, repeat mistakes, and fail in weird ways when tasks stretch across sessions. “Dreaming” is Anthropic’s attempt to make agents compound, not just respond. That is a much bigger commercial promise. (venturebeat.com) ### Where do FDEs fit in? FDE means Forward Deployed Engineer. Anthropic uses the term for engineers who embed with major customers and help ship real AI applications inside the business. So when people talk about commercial FDEs here, they mean Anthropic is not just selling a model endpoint. It is selling the playbook, the integration work, and the hands-on team that makes agents usable in production. (venturebeat.com) ### What is the catch? More vertical integration cuts both ways. Anthropic’s managed-agent push makes life easier for companies that want one vendor to handle memory, orchestration, and runtime. But it also raises lock-in concerns, because the same convenience means Anthropic controls more of the stack. Startups building agent infrastructure may read this two ways — as validation of the market, and as a warning that the model provider wants to own more of it. (anthropic.com) ### Bottom line? The doubled limits are the visible part. The real story is that Anthropic is trying to turn extra compute into longer-running, more dependable, self-improving agents. If that works, Claude stops being just an AI assistant and starts looking much more like enterprise software with initiative. (anthropic.com) (venturebeat.com)