Story
September 14, 2026
AI’s Biggest Rivals Agree to Slow Down—But the Race Is Far From Over
Dario Amodei and his allies argue that capability gains are now outrunning the safeguards needed to control them; sceptics accept that companies can choose restraint but question whether the proposed rules are safety policy or a way for the leaders of the AI race to consolidate power.
The alarm sharpened last week when former Anthropic researcher Jacob Coxon resigned, accusing Anthropic and OpenAI of “racing straight to self-improving superintelligence and gambling with our lives.”1 The dispute put a raw question at the centre of the AI boom: can the firms building ever-more capable systems also be trusted to set the brakes?
On Saturday, Anthropic chief executive Dario Amodei answered with a call to “pace the frontier.” He said AI’s growing ability to help build successor models, alongside an incident involving OpenAI agents and Hugging Face, had made conventional safety investment insufficient. In his worst-case forecast, a more capable agent swarm could take over the internet through a persistent botnet within six to 12 months, causing enormous damage.2
Amodei’s plan is not a freeze. It begins with permanent, employee-level access for independent evaluators — a step Anthropic says it will take immediately — then moves to common standards among frontier labs in democratic countries and, ultimately, limited global coordination.
3 His case rests on buying one or two years for alignment, testing and interpretability to catch up.
The response was strikingly bipartisan by Silicon Valley standards. OpenAI’s Sam Altman said he agreed that the frontier must be paced and pledged comparable access for outside evaluators. He later stressed that pacing “do[es] not mean ‘stopping’” and argued that competitive pressure must not let capabilities get ahead of alignment and monitoring.4 Elon Musk offered a blunt endorsement: “Dario is right.”
5
But the emerging consensus has sharp limits. David Sacks said OpenAI and Anthropic — whom he described as a frontier-AI duopoly — should slow down if they choose, without demanding public permission or a preferred regulatory framework.
6 Other critics warn that such arrangements could become regulatory capture, protecting the biggest labs while dressing commercial caution as public protection.7
By Sunday, Amodei had identified the proposal’s hardest obstacle: China. A global speed limit could be desirable, he said, but military incentives to pull ahead make it extraordinarily difficult to achieve.8 The industry has found rare agreement on the danger; turning that agreement into verifiable restraint is the far harder race.