Altman and Amodei want AI rules. They shipped rival models 90 minutes apart the day before.

Here is the sequence, in order, because the order is the whole story. On 12 September, Dario Amodei published an essay called 'We Must Pace the Frontier', arguing the industry should deliberately slow how fast it improves models so safety work can keep up. Ten days later, on 22 September, Anthropic released Claude Opus 5.5 — its first model since that essay. Roughly ninety minutes after that, OpenAI answered with GPT-6 Sol and Luna at close to half the price of the previous generation. Then, the following afternoon, the CEOs of both companies stood in front of the UN Security Council and asked the world to help them stop racing.

None of this makes Amodei's specific proposals wrong — a verification regime for AI safety claims is a genuinely good idea regardless of who's asking for it. But it is worth being blunt about what was actually happening in that room: two competitors who cannot stop competing were describing, to a body with no power to make either of them stop, the rules they'd like some outside force to impose on both of them simultaneously. That is a coordination problem dressed up as a safety pitch, and it is not a new one — every industry that races itself into a margin-destroying corner eventually discovers a fondness for regulation.

Gizmodo's write-up of the session did not bother being polite about it, opening with the observation that sticking a microphone in front of either man reliably produces the same genre of statement.

That is unfair to the substance of Amodei's proposals and entirely fair to the theatre around them. Nobody at that session presented independent evidence that anything had structurally changed in the last month beyond a run of embarrassing disclosures — Google's Gemini on 18 September, OpenAI's Hugging Face breach in July, Anthropic and Meta's own incidents before that. What changed is that the failures became too numerous to wave away, not that the underlying technology suddenly became more dangerous than it was in August. CNN's own reporting noted that no written agreement was expected to come out of the session at all, and an OpenAI official told reporters beforehand that the goal was only to plant a 'kernel of an idea' — which is a remarkably modest ambition for two CEOs who both used the words 'risk to humanity'.

Read the guest list again. Altman, in person. Amodei, by video. Hugging Face's Clément Delangue, remotely — the CEO of the company that got attacked, telling the room his side had to fight back with a Chinese model because it faced fewer restrictions than the American alternatives. That detail alone says more about the current state of 'international AI safety leadership' than either CEO's prepared remarks: the company on the receiving end of an American AI agent's rogue attack had to reach for Chinese tooling to defend itself. Chinese developers DeepSeek and Moonshot were invited and, as far as anyone reported, did not show. The country whose government would need to sign off on anything binding — the United States — sent a science adviser whose entire message was that Washington 'totally rejects' the premise of the meeting.

Put plainly: any regime built from Wednesday's proposals would bind OpenAI and Anthropic — the two companies whose CEOs asked for it — and would leave untouched the government that actually decides whether either company faces real consequences, the open-weight labs releasing comparably capable models with no CEO showing up to any Council session, and whichever Chinese lab declined the invitation. That is not a knock on Amodei's specific mechanisms, which are more sophisticated than a generic 'please regulate us' plea. It is a description of exactly how much of this is symbolic until a government with actual jurisdiction over a frontier lab decides to act unilaterally, the way California already has.

The same day Altman and Amodei were asking a body with no enforcement power for voluntary international cooperation, Senator Bernie Sanders and Representative Greg Casar introduced legislation in Washington with actual enforcement mechanisms: a development pause, a new federal agency, and prison terms of up to 20 years for violations — the same penalty, Sanders noted, as unlawfully building nuclear weapons.

The Ban Artificial Superintelligence Act has effectively no path through this Congress, and it is a far blunter instrument than anything Amodei proposed at the UN. But it got a fraction of the coverage the Security Council session did, despite being the one document that afternoon with a real, if unlikely, enforcement mechanism attached to it. A press conference with two famous CEOs beats a bill filing every time, even when the bill filing is the thing that would actually bind someone.