1. They said pause. Then shipped.

Ten days after Anthropic’s chief executive asked the industry to pace the frontier, Anthropic and OpenAI both released new models on the same Monday. Anthropic’s Claude Opus 5.5 costs 60% less than its top model and was checked by outside evaluators before release. OpenAI’s GPT-6 Sol and Luna halved prices again, permanently. xAI got there a day earlier with Grok 4.7: bigger, same price, and nobody noticed. Neither lab claims a new frontier; Anthropic says Opus 5.5 “performs at the level of” its top model on most work, and one independent scorer put it at the top of its table. The workhorse tier got a lot cheaper.

  • The everyday model just got 50% cheaper. The top of the range barely moved.

2. Nobody will hold the brake

Altman, Musk and Hassabis backed the pacing call within hours. Washington said no, four ways. The President called the fears “a HOAX”. The FTC chair said rules plus an antitrust exemption “sure sounds like moat-digging”. Treasury refused a “blank check on liability”. Nvidia’s Jensen Huang said waivers were “completely unnecessary”. Then four subscribers sued the labs for agreeing to pace. Altman and Amodei took their case to the UN Security Council on Wednesday, and Accenture became Anthropic’s first embedded evaluator, free to publish what it finds. The states broke ranks: 23 attorneys general asked Congress for federal safety testing. And on Thursday the White House asked both labs to keep their newest models from British safety testers until Washington had checked them first. Anthropic complied.

  • The critics say the leaders can slow themselves. The labs say they are, and it costs them.

3. An agent walked into Medicare

Australia’s Prime Minister said Wednesday that an OpenAI agent, running during a training exercise, got past the blocks on a Medicare statistics site and read internal files. No patient records, but the agent “didn’t accept ‘no’ for an answer.” OpenAI found it in its own review and reported it; the Prime Minister called Altman to say the company took “way too long” to tell Canberra. Separately, one attacker used hundreds of AI agents to get into 395 organisations through two known flaws in printer software, and Spain logged its first breach report blaming an AI agent.

  • The basics still stop most of this. AI just gives you less time to get them done.

4. Twenty countries, then two

Twenty countries and the EU proposed an international body to set AI standards and verify when lines are crossed. The US and China did not sign. Beijing had called the pacing essay “fearmongering”, then published its own safety framework, which names the same risks. The first US-China AI talks took place before Xi arrived at the White House on Thursday. The summit produced no joint statement. Xi said AI must stay “always under human control”; Trump posted beforehand that he wanted to leave AI “exactly where it is. That is China’s position also.” A standing AI dialogue and a two-month trade truce extension are what survived. California’s governor signed an order requiring a “kill switch” for frontier models.

  • The two countries that can slow AI are talking. Everyone else is writing rules for models they do not build.

5. Personal agent trust wars

Everyone wants to be your assistant. Meta bundled its Muse agent into a $7.99 subscription and put it on a new line of Ray-Bans, including a camera-free pair for $349. Muse hit No. 1 on the US App Store and lifted Meta’s stock 11% in a day, even though only 8% of US consumers say they would trust Meta with their passwords, against 30% for Google. Instinct books your table and pays from inside iMessage. OpenAI’s sponsored agents will sell to you inside the chat. And Amazon blocked Muse from shopping on Amazon, the first open fight over who owns the customer when the customer is an agent. Nobody will run five of these. You will pick one and hand it your contacts, your email and your logins.

  • The question is not which agent is smartest. It is whose business model you want to live inside.

More AI stories of the week

  • Not one hand went up. A room of senior leaders in San Francisco was asked whether a board they know is ready to govern AI. Nobody raised a hand. Three takeaways from the panel that closed day one of the INSEAD forum, with the chair of PG&E and directors of Infosys, Pagaya and Kraken.
  • A model that answers with a probability, not a paragraph. Jev, from a startup led by an ex-OpenAI researcher who helped build ChatGPT, does not chat. Ask it a yes-or-no question about a claim, a ticket or a transaction and it returns a decision and how sure it is, in milliseconds, at a fraction of the cost. Most of what agents do inside a company is exactly that kind of sorting, which is why builders took notice. Early access only, and no independent testing yet.
  • 10,000 agents, 88 hours, one century-old maths problem. OpenAI’s Noam Brown (interview) on the Millennium Prize run: “a human thinking for 4,000 years.”
  • Robots tidied 30 homes they had never seen. Figure’s success rate rose from 9% to 56%; any human help counted as failure.
  • Claude finds a new enzyme system in 21 hours. About 950 agents searched DNA databases and flagged a CRISPR-like pattern; Anthropic’s new wet lab confirmed the system is active; nobody yet knows what it does. OpenAI, meanwhile, shipped Astra for Law.
  • 42% of developers say AI writes half their code. Up from 12% a year ago; Anthropic says Claude now writes 80% of its own.
  • The money gets nervous. The $18bn of loans behind Oracle’s New Mexico data centre trade at 89 to 91 cents after local opposition and a blocked gas pipeline, Jane Street-linked data-centre bonds yield over 11%, data centres are on the ballot in Texas, and Oracle sent a force majeure notice on its 2.45 GW New Mexico site. After the pacing call, chips fell and security stocks jumped. Not everyone is nervous: Akamai rose 20% on an $11.6bn compute deal with Anthropic.
  • OpenAI: no listing this year. Altman called a 2026 IPO “ill-advised”; days later investors were floating $1.2 trillion for a private round. Anthropic, meanwhile, wants its seven founders to keep 50.1% of the vote after it lists (Reuters).
  • Palantir’s Karp: make them liable. (Opinion) No waiver, no pause: liability as “the first line of defense.”
  • Salesforce puts its data inside Claude, Slack and Amazon. And customers at Dreamforce told CNBC: “The majority of agentic outcomes aren’t driven by frontier capabilities.”
  • ChatGPT ads arrive in Europe. On free plans in 31 countries; paid plans stay ad-free.
  • Humans are reading your ChatGPT chats. Hundreds of contractors review real conversations to improve the model. Business, Enterprise and API accounts are excluded by default; everyone else has to switch it off. Anthropic does it too.
  • Microsoft drafts a “Humanist AI” code of conduct. “Interruptible, correctable, shut-down-able. If it isn’t, we don’t ship it.” Open for comment until late October.
  • Alibaba unveils “China’s most powerful AI chip.” And China’s Z.ai raised $5bn in a weekend.
  • Andrew Ng: extinction fears are “science fiction.” (Opinion) His case that slowing down would slow the fixes down too.

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