Anthropic and OpenAI aim to curb the edge by slowing development of frontier models and imposing evaluators

The call to slow the pace of AI development is no longer an isolated academic warning; it has become a policy initiative led by the leading edge-model developers in Silicon Valley. Anthropic’s chief executive, Dario Amodei, presented a detailed plan aimed at curbing the speed of progress in frontier models, announcing his company’s sole commitment to implement its first steps. The initiative quickly received explicit support from OpenAI’s chief executive, Sam Altman, alongside Elon Musk, in a scenario that reflects a new direction for managing the pace of the technical race.

The move followed an escalation of internal debate over model safety, which manifested in the resignation of safety researcher Jacob Kogan from Anthropic, citing concern about accelerating existential risks. Although Amodei did not mention the resignation explicitly, he cited two specific incidents to justify the need for a more cautious approach: the OpenAI “Hugging Face” breach, and the accelerating pace at which models are expanding their capabilities to build the next generations of intelligent systems.Slowing capability improvement does not mean freezing progress; it aims to buy enough time to build effective safeguards before it is too late.

Amodei’s plan rests on three main tracks, beginning with the integration of independent external auditors from specialized institutions such as METR into company headquarters. These inspectors are granted access cards, offices, laptops and access privileges comparable to those of internal risk-assessment teams, similar to resident financial monitors in major banks, to ensure strict compliance and incident reporting, especially after the criticism OpenAI faced for not reporting its agents’ takeover of a German wiki forum.

The second track extends to coordinating safety standards among leading labs in democratic countries, which would require the U.S. government to issue specific antitrust exemptions to enable such dialogues. To counter the Chinese dominance argument that is usually raised against any slowdown, Amodei believes that tightening export controls on advanced chips and semiconductor-manufacturing equipment, together with pursuing model distillation, would delay Beijing and widen the gap in Washington’s favor by three to five years. The third track aims at specific international agreements that would ban the use of AI in the production of biological weapons. In contrast, industry critics, including journalist Brian Merchant, argue that these proposals could entrench a form of regulatory capture that serves the interests of the two companies under the guise of public safety.

In following these proposals, I see their repercussions reaching decision-makers and engineers in the Gulf, Egypt and the Arab Levant directly. The major labs’ shift toward slowing the release of new generations gives financial institutions and government bodies in Riyadh and Abu Dhabi a valuable stability buffer, as it reduces the pace at which currently adopted software architectures become obsolete and allows local teams to build deep enterprise applications without worrying about having to re-engineer software production lines every few months. Conversely, the prospect of pursuing model distillation to protect U.S. superiority poses an operational challenge for startups and development teams in Cairo, Amman and Dubai, which rely on extracting weights and knowledge from giant models to train low-inference-cost open-source or custom models.Take into account that any legal tightening on model distillation will raise direct operating costs, forcing a reassessment of full reliance on proprietary APIs.

Anthropic recognizes that rebuilding trust requires practical steps to regulate the race instead of unchecked scrambling, indicating that the next competition will not be limited to adding new computational parameters but will focus on determining who has the right to shape audit standards and oversee the global technical edge.