Sam Altman Rejects Anthropic's Push for Binding AI Rules

Sam Altman and Dario Amodei lead the two labs racing hardest toward the same frontier. This week they told the public two different things about what that race should cost. Altman said the world should accept some harm for AI's benefits. Amodei has spent months arguing the opposite. Frontier models, in his view, need the same binding tests airplanes clear before anyone boards one. Jacob Coxon, who worked inside both labs before he quit, told New York City's Council on Monday that trust in either position still runs short. The three men kept to separate stages this week. Yet their words answer the same question from three different rooms. Both labs spend billions chasing the same models, and both now ask the public to trust their own account of what safety requires.

A Resignation Crosses Into a Council Hearing

Jacob Coxon spent three years on OpenAI's technical staff, then joined Anthropic's pretraining team earlier this year⁵. He resigned Sept. 8. His post warned that systems are racing toward self-improvement faster than anyone has learned to steer them⁵. "On the current path, I think it is more likely than not that humanity loses control to these AIs, and it could end in human extinction," Coxon told New York City's Council on Monday⁴. Alex Turner, a researcher who left Google DeepMind in June, made a related point about where that risk originates. "China is not our only potential adversary. With reasonably high chance, we are racing to build and grow our own adversary here at home, which is misaligned AI," Turner said⁸. Morgan Dwyer, OpenAI's head of policy development and operations, answered from the same hearing room. "None of these levels is remotely acceptable. We should not train models that we cannot make an extremely strong case that we can keep under human control," Dwyer said⁶. Logan Graham, who leads Anthropic's Frontier Red Team, testified at the same hearing. His own team had just published research on a rival's open model. It found the model's safeguards could be stripped for about $4,400 in compute⁹.

Altman Draws His Own Line at Catastrophic Risk

The same Monday, Politico published Sam Altman's clearest statement of where he draws his own line. "We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency," Altman said¹. He narrowed that claim a moment later. "Accept bounded risks, accept risks that we understand in exchange for the benefits, and liberty, agency, all those things. But don't accept, like, the really catastrophic risk," he said¹. Altman had already named a number for that threshold five days earlier. The occasion was an interview about OpenAI's funding plans. "I think it is unacceptable to be taking like a 10% chance of killing everybody by the end of the decade," he told TechCrunch on Sept. 29³. Ten percent, in his own words, sits on the far side of a line he calls unacceptable. Where the acceptable harm stops and the unacceptable kind begins, Altman alone has drawn so far.

Amodei Wants an FAA for Frontier Models

Five months before Altman's podcast interview, Amodei published his own answer. The essay, posted June 10, carried the title "Policy on the AI Exponential." "Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety," Amodei wrote². His proposal names four areas for mandatory review: cybersecurity, biological weapons, loss of control, and automated research that could speed up the other three. Aviation regulators ground a new aircraft design until independent tests clear it. Amodei's essay asks Washington to apply that same standard to frontier training runs. A company's model clears a compute threshold, then a qualified outside party tests it, under Amodei's plan, before the model ships at all. Altman's company and Amodei's both signed a lighter version of that idea on Oct. 1. The occasion was a White House luncheon Trump called "morally binding"⁷. Six executives put their names on one page: Amodei, Altman's president Greg Brockman, Sundar Pichai, Mark Zuckerberg, Elon Musk and Jensen Huang⁷. A one-page pledge and a four-part testing regime are two different sizes of the same idea.

The Council Presses for a Number Over a Principle

Council Speaker Julie Menin tried to pin the panel to a figure rather than a philosophy. "I think a simple yes or no would instill more confidence in the public on a matter as serious as this," Menin said⁶. Every executive at the table, Dwyer among them, answered with a qualifier instead of a number⁶. Dwyer's own answer, read back, sits between Altman's bounded-risk argument and Amodei's binding-test argument. OpenAI's representative declined to name a percentage. Anthropic's own witness, Logan Graham, discussed categories of risk and skipped naming one too⁶. Shane Cahill, Meta's representative, reached for similar language when Menin pressed him directly. "I don't wish to be imprecise in relation to the question of quantifying. I don't have our framework in front of me right now," Cahill said⁶. A city council asked the question both chief executives had already answered from separate podiums. The room where their own employees sat stayed quiet.

What Each Side Would Need to Be True

Altman's wager needs the public to trust a single company's own definition of catastrophic. His framework keeps that definition inside OpenAI alone, and he could close that gap by letting an outside party test where his own red line actually sits. Amodei's wager needs regulators capable of running a technical audit as fast as frontier labs ship models. That capacity, governments have yet to build, and his own essay leaves open who would run it first. Coxon's warning needs other insiders willing to put the same claim on the record. Alex Turner did exactly that, telling the same council the country risks growing its own adversary at home⁸. Each wager rests on a different kind of proof, and each man's proof stops short of settling the argument for the other two.