Altman committed to evaluators, Musk said “Dario is right”, Hassabis and Nadella backed the direction. Rival labs do not normally agree on anything, which suggests the unreleased models are showing something new.

Then the people who could act said no. Trump called AI taking over the world “a HOAX” and argued a “High IQ” president is the only guardrail needed. Beijing called the plan “fear-mongering” and a “Cold War playbook” for keeping China off the best chips. Zuckerberg broke ranks from inside, arguing each lab already has the incentive to pace itself, because nobody wants an agent that ignores instructions, so alignment is a selling point. Scorecard: four labs for a pause, Zuckerberg, Huang, Trump and Beijing against. Acceleration is the default. The European voice came from a finance ministry: France’s Roland Lescure called the slowdown talk “self-interest” and said Europe should speed up its use instead. He has a point. We are not setting the pace of the frontier, we are setting the pace of adoption, and Statistics Estonia put that at 34% of companies using AI, 6.2% using it for a real process or decision.

Here is the part worth keeping. The pause did not happen, but the paperwork moved, all in one week. OpenAI published six reports on its own models misbehaving. Microsoft AI published a code of conduct that outranks its own users. Nvidia and Palantir restricted a model over a data retention policy. So what does this mean for the rest of us. Most of us are not deciding how fast AI gets built, we are deciding what we let it touch this quarter. My simple advice: write down which of your processes an AI tool may run without a human signing off, and which ones it may not. One page. Start with that.

What they are pointing at is recursive self-improvement. Today humans build the models: we design the training, write the tests, decide what ships. Recursive self-improvement is the point where the model does that to itself, and each version is better at improving the next one. The fear is not a model that hates us. It is that the improvement loop stops running at human speed. Hubinger was clear that today’s models are low risk; the worry is that loop starting sooner than planned.

Hold it at arm’s length, though. Some of this is positioning: “too dangerous to release” has been a sales register in this industry for years, and this is months before the US midterms, when AI is an election issue. Not everyone in the field agrees either. What settles it for me is the capex. Anthropic has committed $517B to compute, up from a $120B projection last December. OpenAI is at $750B through 2030. And 1,400 industry leaders, Dario Amodei included, signed a letter asking government to help them slow down, then everybody kept building. A pause is not coming: Anthropic is targeting an IPO in late 2026, OpenAI in Q1 2027, and any deal to pace each other runs into antitrust.

My simple advice: do not spend your Monday on extinction. Spend it on the fact that the people closest to this are telling you it keeps getting sharply better and nobody is holding the brake. That is a planning input, not a panic. The AI capability you are budgeting for next year is a floor, not a ceiling. And the boring things matter more, not less: who may connect an AI tool to real data, who checks the output, who is accountable when it is wrong. You cannot govern superintelligence from a finance team. You can govern what runs inside your own walls, and most of us have not done that yet.

The price spread is the part most people skipped. Gemini 3.8 Flash costs $0.75 in and $3.75 out per million tokens. Astra costs $10 and $50, roughly 2.5x the model it replaces. That is more than a tenfold gap between the cheapest and the most expensive model on this list, for a difference of two points on the index. And the headline price is not the real number anyway. Anthropic claims around 25% cost savings on typical work with Fable 5.1, but Artificial Analysis found a maximum-effort task actually costs 20% more, because 5.1 writes about 1.7x the text. Cost per token tells you almost nothing. Cost per finished task tells you what you will actually pay.

Something else moved this week, and it moved in the opposite direction from safety. Fable 5.1’s filter now steps in 60% less on cybersecurity work and 85% less on basic medical and biology questions, and Anthropic released Mythos 5.1, the same model with fewer guardrails, to screened US researchers. OpenAI rated Astra its first “Critical” cyber risk. And The Information reported that Astra uses a technique called recurrent depth, which loops the model over the same text repeatedly and produces thinking output that is pure math rather than readable reasoning. OpenAI reportedly dialled the loops back so the reasoning still gets written out. Capability went up this week. Visibility into how these models reach an answer went down.

Then there is the supplier question, which arrived from a completely different direction. OpenAI told Cursor it will remove its models by November 12, because SpaceX bought Cursor and a cancellation right opened. Whatever you think of the reason, the mechanics are what matter: a company built a product on someone else’s model, ownership changed hands above its head, and it got about ten weeks’ notice.

My simple advice.Do not pick a permanent winner this week. The leaderboard changed four times in five days and it will change again. Build so you can swap the model underneath without rebuilding the work around it. Measure cost per finished task, not per token, because the two now point in different directions. Keep the evaluation you run on your own real work, because that is the only benchmark that scores what you actually do. And when a model is doing something that matters, know how long it would take you to move off it.