AI enthusiasts scrambled over the weekend after Anthropic blocked Claude subscriptions from powering third-party agent tools such as OpenClaw.

Why it matters: The move underscores a growing tension at the heart of the AI boom: Power users want autonomous agents that run constantly, but AI labs are trying to control costs, capacity and how their models are used.

Driving the news: Anthropic's Boris Cherny announced the change on X late Friday.

• Users can still access Claude models — including Opus, Sonnet and Haiku — through outside agent frameworks.
• But they'll now need to pay via Anthropic's API or a new pay-as-you-go "extra usage" system, rather than relying on flat-rate subscriptions.
• "The $20/month all-you-can-eat buffet just closed," writes AI product manager Aakash Gupta.

Catch up quick: OpenClaw is a popular open-source tool that lets users run autonomous AI agents continuously.

• These agents can burn through far more tokens than typical chatbot use — sometimes running 24/7.
• Claude's $20/month subscriptions had effectively allowed heavy users to access high-cost model usage at a flat price.
• By pushing developers toward API billing and paid add-ons, Anthropic gains tighter control over pricing, rate limits and margins.

Between the lines: Compute is finite, and power users were straining the system.

• Anthropic had already introduced stricter usage limits, including five-hour session caps during peak periods.
• Heavy agent usage can reduce availability for everyday users.

The big picture: Expect more friction like this. Agentic AI is far more expensive to run than chatbots, and someone has to pay.

• These systems can run for hours and take actions across apps, so AI providers end up footing the bill for the compute costs incurred by super users.
• As Anthropic's Cherny put it: "Our subscriptions weren't built for the usage patterns of these third-party tools."
• "Capacity is a resource we manage thoughtfully and we are prioritizing our customers using our products and API," he said.

Yes, but: The shift is also fueling backlash from the open-source community.

• OpenClaw creator Peter Steinberger, who was hired by OpenAI, said he pushed Anthropic to reconsider and shared workarounds after the change.
• Some users are exploring locally run models to avoid usage limits altogether.

Zoom out: The move reflects a broader strategic divide in AI: efficiency versus growth.

• Critics have argued rivals like Sam Altman's OpenAI lack financial discipline, while Anthropic is successfully pitching Wall Street on a more capital-efficient approach.
• A former employee of several different frontier AI labs told Axios that Anthropic emphasized efficiency in how it trains and runs models, while the mindset at OpenAI was that Altman could always raise more capital to support scaling compute.

Bottom line: The faster agents get more capable, the more the business model (rather than the tech) becomes the bottleneck.

A growing number of software developers say AI coding tools are frying their brains.

The big picture: The most popular agentic AI systems have triggered something that looks a lot like addiction among some of tech's highest performers.

Catch up quick: Agentic coding tools like Anthropic's Claude Code, OpenAI's Codex and the open-source tool OpenClaw can write, test and ship software autonomously.

• The developer prompts, watches, reviews and then prompts again.
• It sounds great. Until it isn't.

What they're saying: OpenAI co-founder Andrej Karpathy — coiner of the term "vibe coding" — told the "No Priors" podcast he's been in a "state of AI psychosis" since December, trying to figure out what's possible and "pushing it to the limit."

• Karpathy says his ratio of hand-written to AI-delegated code flipped from 80/20 to 0/100 in December.
• He now spends 16 hours a day issuing commands to agent swarms.
• Karpathy pays a monthly subscription fee and when he has tokens left over near the end of the month he says he "feel[s] extremely nervous" and rushes to exhaust his supply in order to keep up with everyone else.

Y Combinator CEO Garry Tan has called his experience grinding with coding tools "cyber psychosis" and posted in January that he "stayed up 19 hours yesterday and didn't sleep til 5AM."

• In response to a startup founder bragging that his CTO hadn't slept in 36 hours, Tan said: "This is unhealthy by the way (speaking from experience)."

AI developer and blogger Simon Willison, who has 25 years of pre-AI coding experience, said on "Lenny's Podcast": "There is a limit on human cognition, in how much you can hold in your head at one time. And it's very easy to pop that stack at the moment."

• Developers need to know their own limits and figure out responsible ways to prevent burnout, he says. Choosing agentic coding over sleep is "obviously unsustainable."

The intrigue: Work with agentic coding tools is starting to look less like a fun quirk and more like a pathology.

• There are elements of gambling and addiction in the way people are using these tools, Willison said on "Lenny's Podcast."
• "Many of us got hit by the agent coding addiction. It feels good, we barely sleep, we build amazing things," software developer Armin Ronacher wrote in January.

Reality check: Serious coders have always been "locked in" to meet deadlines, and pulling all-nighters is nothing new.

• Elon Musk and his crews have been sleeping at the office and on factory floors for years, but at least they slept.

Zoom in: Quentin Rousseau, CTO and co-founder of the incident management platform Rootly, told Axios he couldn't sleep for months after switching to agentic coding. Eventually he needed a doctor to prescribe sleep medication just to shut his brain off at night.

Keep reading.

• OpenAI CFO Sarah Friar reportedly disagrees with Sam Altman on when the company should IPO. (The Information)
• Here are the states that don't want to house data centers. (Axios)
• OpenAI shuffles its leadership team as OpenAI's head of AGI deployment, Fidji Simo, takes medical leave. (Axios)

With the UCLA Bruins winning the 2026 NCAA women's hoops tournament, the results of this year's Axios AI+ bracket challenge are in. Congrats to Constellation57 W, who had the highest total. (Email us and we will send some Axios swag your way.)

• Also, shout out to my whole family. AJ and I finished tied for fourth place and Harvey (making his prognosticating debut) finished in ninth place, with us all correctly picking UCLA as the eventual winner.
• Claude (33rd place), ChatGPT (40th place) and Gemini (68th place) each had UConn winning.