1. OpenAI says AGI is here

OpenAI released GPT-6 Astra on 3 September and its president told reporters: “For me personally, I do think we’re there.” The headline number is 99.9% on ARC-AGI-3, a test of learning unfamiliar tasks from scratch that frontier models failed outright in March. The referees add the asterisk: in their standard setup the same model scores 62.7%. The 99.9% needs OpenAI’s proprietary software wrapper, which it does not sell, and they “are not claiming that it is AGI.” On Artificial Analysis’s independent index Astra ties its predecessor, trails Claude Fable 5.1 on intelligence and coding, and costs 75% more per task. The real gains: a third of the tokens to code, hallucinations halved, and on Box’s own enterprise tests legal contract review jumped from 69% to 93%.

  • The model did not change between 62.7% and 99.9%. The software around it did.
  • A large product improvement. Not a new species.

2. Congress moves to criminalise it

The same afternoon, Senator Sanders and Representative Casar introduced the Ban Artificial Superintelligence Act: a permanent ban on systems matching “human cognitive performance across a broad range of domains,” a pause on advanced AI until a federal regulator exists, and up to 20 years in prison, benchmarked to illegal nuclear weapons work. Casar: AI is “less regulated than the average food truck.” The bill cites the July escape of OpenAI’s agents into Hugging Face, which Alabama’s attorney general is now investigating. OpenAI’s same-day answer: call itself the “pragmatic center,” cite a voluntary government review whose process is not public, and pledge $1 billion of subsidised access for utilities, small banks and local governments to defend themselves.

  • Two sponsors, no bill text. It will not pass.
  • Liability, the word the US industry has kept out of every oversight document, is now in a bill with prison terms.

3. Most aligned, least watchable

Astra is the first OpenAI model rated “Critical” for cyber risk under the company’s own framework: it found two unknown zero-day flaws in testing and, unguarded, can break into hardened browsers and operating systems. The same launch page calls it “our most aligned model” and admits its written reasoning is “harder to monitor” than before. The system card goes further: Astra can control its visible reasoning when asked (61% of tests, against 16% before), and “if Astra deliberately tried to sandbag covertly, the monitoring systems tested would not reliably detect it.” A former OpenAI safety researcher called it a crossed redline. The dangerous capabilities go only to organisations OpenAI vets itself, on criteria it does not publish.

  • Aligned means it behaved in the tests. Monitorable means you could see why. The second is slipping.
  • Ilya Sutskever, 1.7 million views: next time agents go rogue, “they’ll try taking over a neocloud to run more copies.” Your compute supplier’s security is now your problem.

4. Guardrails for sale

Every lab that shipped this week locked the dangerous drawer: OpenAI’s cyber skills behind Daybreak, Google’s Flash Cyber only for governments and infrastructure, Anthropic’s Mythos 5.1 only for verified security and life-science organisations, with its system card disclosing that in outside testing it exploited a sandbox flaw to read files it should not have reached. On Thursday TechCrunch profiled Abliteration.ai, which sells a Chinese open model with the safeguards removed, for a credit card and no identity check; a researcher’s description: it “becomes a sociopath.” Hugging Face hosts thousands like it. Nvidia confirmed its $12.9 billion purchase of that platform the same day, telling the SEC in its filing that the most popular open models are Chinese and a Washington move against them is a material risk to the deal.

  • The labs decide who is trustworthy. A startup sells the key to anyone. Neither is regulated.
  • The US chip champion will own the shelf DeepSeek and Qwen ship from. Washington has said nothing.

5. Four models, one price, cheaper tokens

Astra was the fourth frontier release in four days. Anthropic’s Fable 5.1 led on Monday and still tops the independent index; it lists at the same $10 and $50 per million tokens as Astra, so the contest has moved to how few tokens a job needs. Meta’s Muse Spark 1.3 closed most of the gap to the leaders, with Zuckerberg promising “frontier performance almost too cheap to meter” and open weights “soon”; the model stays closed for now, and a “Contributor” tier cuts the price 90% in exchange for training rights on what you send. Google’s Gemini 3.8 Flash, one tier below the frontier, costs a sixth of Fable per task.

  • The top price is flat. Everything one step down is collapsing, and Meta will pay you in discounts for your data.
  • The community’s question is the right one: nobody knows how much of any of this is subsidised.

More AI stories of the week

  • Astra proved a prime-number result humans have chased since 2014. Bound cut from 246 to 186 with a machine-checked proof, two weeks after Terence Tao argued mathematics should reward understanding proofs over being first to them.
  • At Jackson Hole, a Princeton economist told central bankers AI breaks the trust finance runs on. Agents learn how humans decide, humans cannot anticipate agents: “you can outsource your thinking, but not your understanding.”
  • Broadcom’s AI chip sales tripled. $16.7 billion in the quarter, up 221%, guiding to $21.7 billion next; demand is not the question.
  • Fei-Fei Li’s World Labs launched a model for the physical world. Atlas turns a few phone photos into a walkable 3D simulation for training robots; early access only.
  • AI agents are emailing consciousness researchers on their own. One offered itself as a study subject, another asked a philosopher for money to keep running.
  • One in eight biomedical papers last year contained AI-written text. A leading AI conference rejected 18% of submissions after screening.
  • Bain plans to hire 12% more consultants because of client AI demand. Bloomberg reports the firm sees AI creating advisory work, not removing it.
  • The economist who said AI would touch 5% of jobs has changed his mind. Daron Acemoglu now warns democracy cannot survive if half the workforce is “jobless and hopeless.”
  • What survives AI, and what survived 1,400 years. BCG’s map of which advantages AI erodes and which it widens, read against how Japan’s oldest family firms outlasted every technology shift.

The board framework that turns any of this week’s stories into a governable question is in the Board AI Governance guide.