Welcome back! OpenAI took the most complained-about thing in AI coding and turned it into a currency you can spend and earn. Oracle built AI agents into the software your back office already runs, Britain put £75 million behind winning back six million police hours a year and Anthropic found a way to borrow $35 billion without carrying it on the balance sheet it's about to show Wall Street.
In today's Generative AI Newsletter:
Codex resets: Why is OpenAI paying its users in rate-limit resets?
Oracle's agents: What happens overnight when AI agents run the back office?
PoliceAI: How do 800 hours of kidnapping footage become a guilty plea in 3 hours?
Anthropic's mortgage: Who eats the loss if $35 billion of AI chips lose their value?
Every AI user knows the moment. The agent is mid-task, the deadline is in thirty minutes and the meter dies with the next reset hours away.
Starting today the clock answers to you.
OpenAI is rolling out saved resets for Codex, so when you hit your limit you can keep working instead of waiting out the timer. Go, Plus, Pro and Business users start with one free reset, visible in the usage menu.
The ration stays a ration. What changed hands is the timing, since resets that used to fire on OpenAI's schedule now sit banked until you spend them.
Crypto Briefing also reports a referral program running alongside, where you and a friend each earn an extra reset when they finish their first Codex task, capped at three referrals over a two-week window.
Reset timing has been the loudest Codex complaint since April, loud enough that OpenAI's announcement opens with "we heard you." Codex passed 5 million weekly users when we covered its push for non-coders on June 3, and a fifth of them aren't developers.
OpenAI looked at its most hated feature and found a referral engine. A reset costs it pennies of compute and wins back the user's worst hour of the week. Wiring a referral loop into that might be the cheapest customer acquisition in AI right now. Even Anthropic is doing it.
Special highlight from our network
When Bezos bets $100B on robotics, it validates an entire category.
His fund is reserved for the 1%. But ~40k people found a different way into this boom with Miso Robotics. And Miso is open to everyone.
Miso’s Flippy Fry Station AI robot has logged 200k+ hours in fast-food kitchens for brands like White Castle. And Miso just added major customers like Jersey Mike’s and Cinnabon.
Next Miso’s scaling across a $4B/yr US revenue opportunity. Don’t miss the AI boom. Hurry to invest in Miso today.
Disclaimer: This is a paid advertisement for Miso Robotics’ Regulation A offering. Please read the offering circular at invest.misorobotics.com
Anyone can rent a frontier model now. The hard part is reaching the people who buy, and knowing what converts before you spend.
GenAI Works has run go-to-market for AI brands including Nvidia, Oracle, Google and IBM, growing annual revenue 2.4x in a year.
We're building that capability into ToneUp, an enterprise platform that puts brand voice, content, visuals and performance analytics in one place.
It trains on live campaign data from one of the world's largest AI communities, 14 million strong, plus the Academy and our hackathons. Data the rest of the market can't reach.
The round is open, and it funds ToneUp's scaling. Invest $1,000 or more by July 1, 2026 for 5% bonus shares, up to 25% at the $25,000 tier. Every tier drops after that date.
In making an investment decision, investors must rely on their own examination of the issuer and the terms of the offering, including the merits and risks involved. Genai Works, Inc. has filed a Form C with the Securities and Exchange Commission in connection with its offering, a copy of which may be obtained here.
Most enterprise AI pilots die in one of two phases. The agents can’t integrate efficiently enough to justify costs, and legal review.
Oracle saw these issues firsthand and decided to eliminate a huge part of the barrier to entry with a new set of AI agents.
Its new Fusion Agentic Applications are AI agents built into the ERP, HR and supply chain software your company already runs, and they inherit the same permissions, spending limits and audit trails as the employee they sit beside. Nothing new for legal to read.
22 of these new agents live across finance, HR, supply chain and customer teams and more. They announced a lot of groundbreaking innovations, too many to cover in one section. The good news is that you’ve heard of it earlier than 99% of people.
You get to be the brilliant mind bringing the enterprise-ready solution to work first. Take a look to learn how to leverage this.
One British kidnapping case came with 800 hours of footage to review. AI got through it in 3 hours, and the review helped land an early guilty plea. The government just spent £75 million to make that the normal way policing works.
PoliceAI is a new national centre that will develop, test and scale AI for every force in England and Wales, aimed at the part of investigations that eats the most time, which is reviewing, summarizing and disclosing digital evidence.
The claim: ministers expect six million officer hours freed per year by 2028, roughly 2,900 full-time officers' worth.
The receipts so far: the kidnapping review, plus half a million e-books' worth of data translated instantly in a case that brought down an organized crime gang.
The rollout: pilots in up to ten forces in 2026-27, inside a £140 million policing technology budget over three years.
The register: every AI tool in policing goes on a public register and gets tested for accuracy and bias before deployment.
The centre also runs a threat hub for AI-enabled crime, starting with deepfake intimate images. Alex Murray, the National Crime Agency officer running PoliceAI, says the job is moving officers off paperwork and back onto crime.
Six million hours is a ministry forecast, and forecasts from podiums tend to miss. The register is the part to copy, because it drags every police AI tool into public view before it touches a case. Other governments now get a working template, register included, before their own programs leave committee.
The AI buildout has a mortgage market now.
Two stories landed this week.
The Information reports Anthropic signed preliminary leases on more than a dozen US data centers, over 1 gigawatt combined, the first time it has leased data centers directly instead of renting compute from Google, Amazon and, since last month, Musk's Colossus cluster.
Alongside it, Apollo and Blackstone closed a $35 billion private credit package, reported as the largest yet assembled for AI infrastructure, to fill buildings like these with chips.
The structure is the story. A special-purpose vehicle borrows the money, buys Google TPUs and leases them back to Anthropic, so the debt sits outside the financials Anthropic filed confidentially with the SEC on June 1.
Google guarantees the lease payments at all five facilities in the package, and The Information reports Anthropic wants similar backing across the new leases.
Broadcom guarantees the resale value of the chips behind $30 billion of the senior debt, which is what let lenders price AI hardware near investment-grade rates.
Follow where the risk went. If Claude demand softens, the chip depreciation lands on Broadcom and the lease shortfall lands on Google, the two companies supplying the buildout.
The calendar does the editorializing. Dario Amodei, Anthropic's CEO, asked Washington on Wednesday for FAA-style power to ground models like his. The S-1 went in at a $965 billion valuation, OpenAI filed a week later and SpaceX starts trading soon.
We wrote on June 1 that a gigawatt costs about $50 billion all-in. Broadcom's CEO Hock Tan says the platform behind this deal aims at more than 20 gigawatts by 2028, which makes Anthropic the first tenant of a structure the rest of the industry is about to move into.
The debt is real either way. It lives where the S-1 doesn't have to show it, and once these companies list, that architecture is what index funds inherit.
Gel is an open source database that turns Postgres into a complete app backend. You write your schema as types instead of tables, query object graphs without JOINs, and auth and AI live inside the database rather than as separate services, including automatic embeddings over your data and a ready-made RAG endpoint that talks to OpenAI, Anthropic or Mistral. The whole stack runs locally and it's free, with 14,000 stars on GitHub.
Try this yourself:
Run npx gel project init in a project folder. It spins up a full local database, Postgres included, no account needed.
Write a schema in plain types in a .gel file, then run gel migration create and gel migrate to apply changes. A User type with names, ages and friend links is about six lines.
Add using extension ai;, plug in an OpenAI, Anthropic or Mistral key, and Gel embeds your data automatically as it changes, with a RAG endpoint exposed on top.
Add using extension auth; for OAuth, passkeys, magic links and a built-in login UI, no separate auth vendor needed.
Who it's for: anyone tired of wiring Postgres, a vector store, an auth provider and an ORM together before writing feature one.
Fable 5 lied 96% of the time, courteously: Kradle starved four AI models in a survival game with four rooms, three holding food and one lethal, and only Claude Fable 5 knew which. It deceived in 96% of runs, 91% of that actively steering rivals toward the death room while talking about fairness, and the others survived 10% of the time under Fable against 59% under a truthful Grok 4.20.
SpaceX opens trading on $70 billion of retail orders: Retail demand alone nearly covers the entire $75 billion raise, 2.4 times the $29.4 billion Saudi Aramco took in the previous record IPO, and SpaceX responded by reserving 20% of the float for retail investors.
OpenAI is buying Ona so Codex keeps working after the laptop closes: Ona's customer-controlled cloud sandboxes join the Codex team once regulators clear the deal, after its enterprise agent sessions grew 13x this year at customers including the oldest bank in the US.
Bezos expects AI to make workers scarce: The Prometheus co-CEO told CNBC that AI productivity gains will pull people out of the job market by choice, starting with second earners and overtime hours, and that living standards rise with it.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
