Welcome back to artificialintelligence.ai — the newsletter that explains AI like you're a smart human, not a machine learning engineer.
This week was one of the loudest in AI history. A new OpenAI model that clicks around your computer by itself. A $12.9 billion acquisition that changes who owns the plumbing of open-source AI. And a bill in the US Senate that would make building "superintelligence" a crime punishable by 20 years in prison.
In today's issue:
🚀 OpenAI launches GPT-6 Astra and says the AGI era has started
💰 Nvidia buys Hugging Face for $12.9B
⚖️ Sanders and Casar move to outlaw superintelligent AI
⚡ Google's Gemini 3.8 Flash undercuts everyone on price
📌 Quick hits: ChatGPT ads hit $1B, Claude lands in Salesforce, Tesla's Cybercab hits Austin
Reading time: 5 minutes.
🚀 OpenAI releases GPT-6 "Astra" — and uses the AGI word
OpenAI released GPT-6 Astra on September 3, its most capable model to date. The headline feature isn't better writing or better math. It's computer use: Astra can operate a computer the way a person does — opening webpages, filling in forms, working through spreadsheets — except much faster.
OpenAI President Greg Brockman went further than the company usually does: "It's not unreasonable to feel that we are now in the AGI era."
THE BREAKDOWN:
ARC-AGI-3 (a reasoning test designed to be hard for AI): Astra scored 98.6%. The previous model, GPT-5.6 Sol, scored 7.8%. That is not a typo.
Cybersecurity: on OpenAI's ExploitGym test it scored 100%, up from 78.5%.
The release was delayed so OpenAI could add safeguards. Astra crossed the company's own "critical cybersecurity capability threshold," which triggered extra monitoring, isolated training environments, and a government review under a voluntary safety framework.
Availability: it went first to enterprise customers in OpenAI's cybersecurity-focused Daybreak program. Plus, Pro, Enterprise, API and AWS access follow "in the coming days." Regular users get a restricted version that refuses advanced hacking tasks.
WHY IT MATTERS: Until now, AI has mostly been something you talk to. A model that reliably drives a computer is something you delegate to. That's the difference between a tool that helps you do your admin and a tool that does your admin. If Astra holds up outside benchmarks, the "AI takes over the boring parts of your job" story stops being a prediction.
The honest caveat: benchmark jumps this dramatic are usually part real capability, part the benchmark being a good fit for how the model was trained. Watch what people actually get done with it over the next month.
💰 Nvidia buys Hugging Face for $12.9 billion
Nvidia confirmed on September 3 that it is acquiring Hugging Face for $12.93 billion.
If you haven't heard of Hugging Face: think of it as GitHub for AI models. It's where the world's open AI models live and get shared.
THE BREAKDOWN:
What Nvidia is buying: 3 million AI models, 1 million apps, 500,000 datasets, and 18 million developers.
Hugging Face turned down a $500 million offer from Nvidia last year. It makes around $150M a year in revenue. That's roughly 86x revenue — a price you pay for strategic position, not for cash flow.
Jensen Huang's promise: "Hugging Face will remain an open platform for the entire AI ecosystem." Nvidia specifically committed that its own chips will not be required to build or deploy through the platform.
WHY IT MATTERS: Nvidia already sells the chips almost every AI model is trained on. Now it owns the place where those models are distributed. Even with the open-platform promise, the company that sells the shovels just bought the map to the gold mine. Expect regulators — and every AMD customer on the platform — to watch closely.
📎 Sources: TechCrunch · CNBC
⚖️ A US bill would ban superintelligence — and jail people for building it
Also on September 3, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act.
THE BREAKDOWN:
A permanent ban on developing AI that surpasses human intelligence or could overthrow a government.
A temporary pause on advanced AI development until federal safety rules exist.
A new cabinet-level agency to regulate and enforce AI safety.
Penalties: corporate death penalty for companies, up to 20 years in prison for individuals — deliberately modelled on the penalties for illegal nuclear weapons development.
Sanders: "The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs." Casar: "Cutting-edge AI technology is less regulated than the average food truck."
WHY IT MATTERS: Realistically, this bill will not pass in its current form. But its timing — the same day OpenAI said the AGI era may have begun — makes it a marker. The political conversation has moved from "should we label AI content?" to "should some AI be illegal?" That shift affects every AI company's risk calculus, whether or not this specific text becomes law.
📎 Sources: Sanders Senate office · Axios
⚡ Google's Gemini 3.8 Flash: the price war continues
Google shipped Gemini 3.8 Flash on September 2 — three weeks after its last release.
THE BREAKDOWN:
Scores 59 on the Artificial Analysis Intelligence Index at high reasoning. That's level with GPT-5.6 Sol and Grok 4.6, below Claude Fable 5.1 (66) and Claude Opus 5 (63).
Pricing: $0.75 per million input tokens, $3.75 output — roughly 6x cheaper than Claude Fable 5.1, and per Artificial Analysis the cheapest model at its intelligence level. Note: that's an introductory rate. The price doubles in 2027.
A separate Gemini 3.8 Flash Cyber variant went to governments, critical infrastructure and software maintainers for finding and fixing vulnerabilities.
WHY IT MATTERS: Frontier intelligence is getting expensive; good enough intelligence is getting close to free. If you're building on AI, the question is no longer "which model is smartest" but "what's the cheapest model that clears my bar." For most tasks, that bar dropped again this week.
📎 Source: The Register
📌 Quick hits
ChatGPT ads passed $1B. OpenAI says its advertising business hit a $1 billion annualized run rate in under 200 days, and self-service ad buying opened across India, Europe, the Middle East and North Africa — now 40+ countries. OpenAI's line: ads are what keep ChatGPT free for over a billion weekly users, they're clearly labelled, and they don't influence answers. Source
Claude moves into Salesforce. Anthropic and Salesforce announced Claudeforce on August 26: Claude powers Agentforce's reasoning engine, becomes the default model across Slack, and gets a Salesforce plugin with 37 prebuilt sales skills. Open beta expected this month. Source
Tesla's Cybercab hit the street. Tesla launched its steering-wheel-free two-seat robotaxi in Austin on September 3, starting with invite-only rides downtown. Source
SpaceX closed the Cursor deal. SpaceX completed its record $60 billion acquisition of Anysphere, maker of the AI coding tool Cursor. Source
🧠 What this actually means for you
Three practical takeaways from a very loud week:
Start thinking in tasks, not chats. If a model can operate a computer, the useful question changes from "what should I ask it?" to "what whole job should I hand over?" Pick one repetitive process you do weekly and write it down as steps. That's your first delegation candidate.
Don't pay frontier prices for non-frontier work. Summarising, drafting, classifying, extracting — cheap models now handle these. Save the expensive models for the hard 10%.
Assume the rules will change. Whatever you're building on AI, keep it portable across models and providers. This week showed that ownership, pricing and legality can all move in a single day.
💬 One question
Would you let an AI operate your computer unsupervised?
Yes, already do · Only with approval on each step · Absolutely not
Hit reply and tell us why — we read every response and feature the best ones.
That's it for today. Forward this to one person who keeps saying they "should probably learn AI."
— The artificialintelligence.ai team
