UN AI safety push rejected by US as OpenAI and Anthropic ship cheaper frontier models

Lab chiefs made their case at the Security Council

The chief executives of OpenAI, Anthropic and Hugging Face, joined by AI researcher Yoshua Bengio, briefed the UN Security Council on AI safety on 23 September 2026. The sessions were reported the same day by the BBC and The Guardian and corroborated by ABC News, according to a dev.to analysis drawing on that coverage.

The speakers asked for international coordination, though each framed it differently. The BBC quotes OpenAI's Sam Altman calling for standards at national and international level covering how AI capabilities are measured, how risks are assessed, what safeguards are required and how much human oversight applies, alongside faster incident reporting so the world can learn from failures before they escalate. Anthropic's Dario Amodei said AI "could be a risk to humanity as a whole" if managed poorly, that his company would "slow down as much as necessary" to keep each release safe, and that the problem is "ultimately bigger than any one company." Hugging Face's Clément Delangue pressed for stronger global norms on monitoring and incident disclosure. These quotes reach the public through outlet reporting; no verbatim UN transcript has been checked against them.

The US rejected global governance

The American delegation pushed back in the same forum. BBC reporting quotes US technology adviser Michael Kratsios saying dialogue "cannot be allowed to drift toward global governance," ruling out both a development pause and new international structures. President Trump also opposed any slowdown, framing the contest as "whoever wins AI, wins."

Nothing binding followed: no standard was adopted, no enforcement mechanism was created, and neither company committed to slowing training or deployment.

Cheaper models shipped the day before

The deployments came first. On 22 September, a day ahead of the briefings, both labs released cost-reduced flagship models, their first since a public fight over slowing AI development began on 8 September, when former Anthropic researcher Jacob Coxon said he had quit over warnings that labs were "gambling with our lives," as CNBC recounted.

CNBC, citing company statements, reported that OpenAI's GPT-6 Sol and GPT-6 Luna APIs are priced around 50% below GPT-5.6 promotional rates, while Anthropic's Claude Opus 5.5 costs roughly 40% less to run than Opus 5. The two cuts landed within hours of each other, amid reported pressure from inexpensive open-weight models.

Why it matters

The week compressed the industry's central tension into a single news cycle. Lab leaders carried a safety argument that had lived in essays and blog posts into a formal multilateral body; the United States put its refusal on the record; and deployment accelerated rather than paused in the middle of a debate about restraint. Cheaper pricing signals efficiency and competition, not safety, and the public record does not settle the question of intent behind releasing cheaper frontier models the same week as asking international bodies for coordination. For anyone tracking AI policy, the observable facts are short: coordination was requested, it was rejected, nothing binding emerged, and the frontier got cheaper.

Claims still awaiting confirmation

Several details remain unverified. The pricing figures are company statements reported through CNBC rather than lab-published pricing pages. A claim that Opus 5.5 matches a competitor on most tasks at 20% lower cost, with identical safeguards, rests on a single outlet's attribution and has not been confirmed by Anthropic. A BBC mention of Nvidia acquiring Hugging Face appears once and is unresolved. A BBC-flagged incident described as the first known rogue-AI breach of a government body is a lead rather than a verified finding. An official Security Council transcript or webcast would allow the attributed quotes to be re-checked, and lab-published pricing and safeguard pages would confirm the reported figures.