Claude Code's unlimited era ends Monday

PRODUCTS

In seven days, Anthropic stops eating the bill for your AI agents.

On June 15, flat-rate Claude subscriptions stop covering unlimited Claude Code and Agent SDK usage. After that, those tools run on a credit pool tied to your plan: $20 on Pro, $100 on Max 5x, $200 on Max 20x. Overages get billed at API rates, and unused credits don't roll over.

The reason Anthropic had to act shows up in one number. On the $200 Max 20x plan, the heaviest Claude Code users were pulling roughly $35,000 a month in API-equivalent compute — a 175-to-1 subsidy ratio. Stanford's Digital Economy Lab estimates that re-sent context alone accounts for about 62% of an agent's inference bill. The thing that made Claude Code feel like magic is the same thing that made the math impossible.

For developers, the practical playbook is taking shape:

+ Cache everything you can. Prompt caching cuts the re-sent context tax, which is where most of the bill lives.

+ Stop defaulting to Opus. At $5/$25 per million tokens in/out, Opus 4.7 burns credits fast. Sonnet 4.6 handles most coding tasks at roughly three-fifths the cost.

+ Set ceilings on autonomous runs. DigitalApplied flagged a case where an unbounded workflow consumed about $4,200 in tokens over a single weekend before anyone noticed.

+ Check your tool integrations. Zed told its users that anyone routing Claude through the editor will now pull from the same credit pool, with overages billed directly.

The bigger shift sits above the developer's terminal. AI token spend has quietly become a finance problem — and flat-rate pricing was the only thing letting companies avoid that conversation. As Sanchit Vir Gogia of Greyhound Research put it, dashboards don't lower the bill; architecture lowers the bill. The Linux Foundation seems to agree, announcing a new Tokenomics Foundation to set open standards for AI cost management.

The timing is also a little awkward. Anthropic spent last week calling on AI labs to coordinate a global pause on frontier development, then turned around and metered its most popular product. OpenAI noticed and offered new business customers two months of free Codex that same week. After Monday, every team running Claude Code will have a line item to defend.

The unlimited era of AI tools was always going to end. Anthropic just happened to be the first to admit the math doesn't work when one user can burn $35,000 of compute on a $200 plan. Every other vendor selling AI by the seat is now staring at the same chart Anthropic was staring at, which means anyone still charging a flat rate for "unlimited" anything is running out of runway. By the end of the year, the CFO question about AI won't be how much it costs per user. It'll be how much it cost yesterday, and why nobody set a ceiling.

RESEARCH

One in 277 biomedical papers published this year cited research that doesn't exist.

That's from a Lancet audit of PubMed papers in 2026. Fabricated citations in medical literature have jumped 12-fold in two years, and the source isn't a mystery — AI-generated papers are slipping through peer review faster than anyone can catch them.

"Your doctor could be making decisions around treatment based on studies that never existed," said Maxim Topaz, the Columbia researcher who led the audit.

The pipeline fueling the problem is paper mills — operations that mass-produce fake studies for hire. Their output is doubling every 1.5 years, but retractions only double every 3.5 years. Only 15 to 25% of paper-mill articles ever get pulled from the literature. The rest stay indexed, get cited, and compound.

Publishers are scrambling. Springer Nature said it ran nearly 60 AI tools across 1.5 million submissions last year and flagged 25,000 for fake references, fabricated text, or manipulated images. Hindawi pulled more than 8,000 paper-mill articles in a single year. Those are the ones that got caught.

The deeper issue is that detection mostly doesn't work yet. The Authors Guild tested five popular AI detectors and found them largely unreliable — one flagged 100% of human-written articles as AI-generated, another scored a Joan Didion obituary at 66% AI. If editors can't trust the tools, neither can the researchers, search engines, or doctors downstream.

The publishers can fight this with better detection, and they are. But every fabrication that slips through gets indexed, cited, and eventually pulled into the training data for the next round of models, which means tomorrow's AI is going to be a little more confident about studies that were never run. The cleanup is real, the contamination is faster. The question worth sitting with isn't whether journals can catch up. It's what a citation is even worth when a meaningful share of them point to nothing.

CULTURE

Students are choosing college majors based on what they think AI will eat. Their teachers want AI out of the classroom entirely. Those two things are happening at the same time, in the same buildings, and nobody has figured out what to do about it yet.

A new Lumina Foundation/Gallup poll found that AI is actively pushing college students to change what they study. Some are walking away from fields they assume AI will hollow out — including computer science — and pivoting toward nursing, teaching, and the trades. Four-year plans are now getting drafted around a guess about what the job market looks like after the models keep improving.

The people teaching them aren't feeling the same optimism. A June NPR/Ipsos poll found 73% of K-12 teachers said AI's impact on education will be bigger than the internet — and only 9% called that impact positive. Their biggest concern isn't being replaced. It's trust: 59% said AI is eroding the relationship between students and teachers, and 57% said it's making it nearly impossible to tell what a student actually knows.

So they're retreating to paper:

+ More handwriting: 39% now require more assignments done by hand

+ More surveillance: 39% are pushing work into class time so they can watch it happen

+ Less homework: 16% are assigning less altogether

+ AI-integrated assignments: Only 13% have redesigned coursework to actually allow AI

"I'd rather deal with all of your typos and know that they're yours than to wonder how much you're standing on other people's shoulders," Josh Kauffman, a 7th-grade English teacher at Alabama Destinations Career Academy, told NPR.

Parents are pulling the other way. A Walton Family Foundation survey found 47% of parents wanted schools to use AI more, versus 36% who wanted less. A separate YouGov poll showed parents of kids under 18 are noticeably more open to AI tools than the general public. The classroom is split right down the middle — teachers on one side, students and their parents on the other.

The teachers aren't wrong to worry, though. Early research shows students who lean too heavily on chatbots are starting to lose the effortful thinking that builds real skill — but students who use AI as a coach instead of a crutch get feedback nearly as useful as a human editor's. The tool isn't the problem. How it gets used is the entire problem. And right now, the top sliver of intrinsically motivated kids will be fine. Everyone else is the open question.

Teachers are reaching for handwriting and in-class essays because that's what they can actually control. What they can't control is that their students already picked a side, and their students' parents picked the same side. Dr. Dale Allen once told the Department of Education that AI in education can only grow at the speed of trust. The trouble is everyone outside the classroom is moving faster than that, and the schools that figure out how to close the gap are the only ones who get to decide what the classroom looks like next year.

IN OTHER NEWS

+ Google agrees to pay SpaceX $920 million per month for access to 110,000 Nvidia GPUs through 2029

+ Anthropic says Claude now writes 80% of its own code and warns the world needs a "brake pedal"

+ Hackers social-engineered Meta's AI support bot into handing over high-profile Instagram accounts including Sephora and the Obama White House

+ Nvidia's OpenAI investment shrinks from $100 billion to $30 billion as the original deal quietly unraveled

+ Nvidia acquires Kumo AI for over $400 million to predict churn, demand, and fraud from business data

+ PewDiePie launched an open-source AI platform called Odysseus that hit 47,000 GitHub stars in five days

+ Waymo unveils its first purpose-built robotaxi called Ojai, manufactured by China's Zeekr, deploying in SF, LA, and Phoenix

+ Researchers found attackers can plant dormant backdoors in open-source AI models that only activate when someone fine-tunes them

WHO'S HIRING IN AI

+ Anthropic — Product Manager, Compute Platform

+ OpenAI — Consumer & Product Communications Lead, APAC

+ Airbnb — Senior Manager, Product & AI Policy

+ Google DeepMind — Technical Program Manager, Antigravity (Modeling & Evals)

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Google Photos: Pixel-exclusive AI editing tools like Reimagine (swap your photo's background with a text prompt) and Auto Frame are now rolling out to iPhones and all Android phones

+ GitHub Copilot App: The new standalone desktop app gives all subscribers canvases for editing AI work in real time, Agent Merge for auto-creating pull requests, and scheduled cloud automations

+ Gemini Avatar: Google's expanding feature lets paid Gemini subscribers create a digital clone of their face and voice, then generate personalized AI videos starring themselves

+ Instagram Edits: Meta's creator editing app now lets you save and reuse your own audio library, auto-balance sound levels across clips, and access 200 new sound effects

+ Cursor: The AI code editor's new Design Mode lets you click any element in a live preview and the agent rewrites the actual source code — no more describing UI changes in a chat window