Welcome back! Anthropic is turning Claude subscriptions into something that looks a lot more like a prepaid utility bill. Starting June 15, developers using Claude through third-party tools will pay API pricing. It lands at a weird moment too because GPT-5.5 has finally made OpenAI feel competitive again in coding. We also have OpenAI floating the idea of China helping govern global AI safety, Meta promising it cannot read your private AI chats and Cisco laying off thousands of employees while Wall Street celebrates.
In today's Generative AI Newsletter:
Anthropic's meter move: Why are Claude power users suddenly getting rationed?
OpenAI's global pitch: Why does OpenAI want China at the AI governance table now?
Meta's privacy promise: Can Meta really sell people on private AI chats?
Cisco's AI layoffs: Why did investors reward one tech layoff and punish another?
Starting June 15, your paid Claude plan splits in two. Use it inside Anthropic's own apps (the website, Claude Code in its terminal app or IDE, Cowork) and your subscription works the way it always has.
Use it anywhere else and you get a separate pot, $20 to $200 a month depending on your plan. When the pot empties, you can switch on pay-as-you-go billing to keep going charged at standard API rates.
This affects you if you use:
Conductor
Zed
Jean
T3 Code
claude -p running inside your CI
Agent SDK called from third-party apps
Claude Code GitHub Actions
Any script that calls Claude Code from outside Anthropic's own apps
Reads like Anthropic pulled the plug on programmatic use across the board, and you start eyeing OpenAI a little harder now that GPT-5.5 has shrunk the gap in coding performance.
This reverses the April 4 ban Anthropic slapped on after Pro and Max users started burning thousands in API tokens through tools like OpenClaw.
The new policy means you should just use the Anthropic front-end for cost efficiency. Everything else, including the way most power users actually run claude -p day to day, now goes on the meter.
By the time it kicks in, running a third-party agent stack is about to cost close to what it actually costs Anthropic to run. What do you think about this new update, and are you switching to Codex after this?
Special highlight from our network
Anthropic ships something new almost every week. Cowork, Skills, Connectors, Design. Most people are still using Claude like it's just a chatbot.
The first Claude-a-thon is a 2-day live workshop on what Claude can actually do in 2026. You'll run real research, build artifacts and dashboards, design full presentations, and set up Connectors that automate things like job searches and inbox triage.
Saturday and Sunday, 10 AM to 7 PM EST. Free for the next 48 hours, 1,000 seats only.
A lot of people try Claude, get mediocre output and end up fixing it manually. The issue is usually the setup. One Claude in a chat won't scale past simple tasks. Agents change that.
There are three types of agents:
Normal agent: one Claude in a loop. Best for simple, sequential work.
Sub-agents: a parent spawns scoped child agents that run in parallel.
Agent teams: multiple agents on a shared goal, coordinated by an orchestrator.
Marc Daou shows when to use each and how to scope them.
Chris Lehane, OpenAI's VP of global affairs, told reporters Wednesday that the US should build a global AI safety body and let China join it.
He compared the idea to the International Atomic Energy Agency (IAEA), the UN-affiliated body that writes nuclear safety rules and counts China as a member.
He said this hours before President Trump's scheduled meeting with Xi Jinping in Beijing, on a delegation that included Nvidia CEO Jensen Huang and Elon Musk.
Lehane wants the new body to plug together the Commerce Department's Center for AI Standards and Innovation and the AI safety institutes other countries are building.
Read between the lines. Strip the IAEA branding off and what OpenAI did was raise its hand for a seat at whatever multilateral body forms next.
The company has spent two years arguing for export controls and against Chinese model access. Pitching a Chinese seat at the table when Trump flies to Beijing looks like positioning for whatever comes after.
Watch whether the White House readout mentions any of it at all.
Meta employees were just protesting the company’s AI tracking and privacy concerns, and at the same time, the company is announcing it’s not going to read your private AI chats.
They’re rolling out Incognito Chat with Meta AI on WhatsApp and the standalone Meta AI app over the coming months.
The promise is that your messages get processed in an environment so secure that even Meta cannot peek, they are not saved by default, and they vanish when you close the chat or lock the phone.
Under the hood, Incognito Chat runs on WhatsApp's Private Processing infrastructure, the same setup Meta built last year to add AI features without breaking end-to-end encryption.
The timing makes the motivation clear. Last month, Reuters quoted lawyers warning that AI chat transcripts can show up in litigation.
Trusting Meta with sensitive AI conversations is a heavy lift. This is the company that leaked Cambridge Analytica data on 87 million users, paid a $5 billion FTC fine for it, and built its core business on harvesting every contact, click and like its apps could capture. What do you think about their announcement?
Cisco, one of the bigger H-1B visa sponsors in US tech, said Wednesday it would lay off nearly 4,000 employees and lift its full-year AI order forecast from $5 billion to $9 billion.
The stock jumped 16% after hours.
Cloudflare cut 1,100 jobs last Thursday (around 20% of staff), beat earnings and lost 24% of its market cap. Makes you wonder why the market reacted differently, is it because of the percentage of employees laid off, or brand reputation?
Cisco trimmed under 5% of headcount while Cloudflare's cuts hit a fifth of the company as gross margins slid from 76% to 71%.
Investors punished one company for cutting amid margin compression and rewarded the other for cutting amid record AI demand. AI is now the cleanest excuse for retiring American roles and the cycle keeps repeating every week.
Juicebox is the recruiting tool that turns a single prompt into a candidate pipeline. PeopleGPT, its core search, hits 800M+ profiles across 60+ data sources from one sentence describing who you want to hire.
It evaluates up to 5,000 profiles per search to surface only the actual fits, then runs autonomous outreach agents that send multi-step sequences and reply 24/7.
Try this yourself:
Sign up at juicebox.ai/signup.
Type a description of your ideal candidate in plain English ("senior backend engineer in Berlin with payments experience").
Browse the live result set, then ask Juicebox to narrow by skill, tenure or location.
Switch on automated outreach to send dynamic multi-step sequences.
For founders or solo recruiters, set up an Agent at Juicebox source, screen and message candidates around the clock.
Claude for Small Business: A new Cowork toggle that wires Claude into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365, plus 15 ready-to-run agent tasks.
Apple opens the door: Apple is designing a system to allow AI agents and vibe-coding apps inside the App Store, with details expected at WWDC on June 8.
Grok Skills: Elon Musk said Grok now supports Skills in the Grok Web client, accessible by typing /.
AI eye scan reads early disease: A professor published an AI eye scan that identifies dementia and Parkinson's in minutes, years before symptoms appear.
Bitcoin recovered with Claude: An X user named cprkrn says he recovered 5 BTC from a wallet locked for 11 years by uploading old college files into Claude.
Learn more about AI from the experts building it
📸 Check our Instagram for AI news in 30 seconds
🐝 Check our Atlas for weekly expert insights
📺 Check our YouTube to learn from the people building AI
🐦 Check our X for AI news as it lands
🧠 Learn how to build your next AI application
