OpenAI is backing the use of independent safety evaluators with deep internal access, bringing it closer to Anthropic’s proposal for more controlled frontier AI development. Sam Altman said OpenAI agrees on the need to “pace the frontier” and has been actively discussing the issue internally in recent weeks.
Altman said the company would give external evaluators access comparable to that available to employees, allowing them to assess safety practices from inside the lab rather than relying only on documentation released after a model launch. The Next Web reported that Altman’s comments came a day after Anthropic CEO Dario Amodei outlined his own plan.
pace the frontier
Sam Altman
Anthropic’s proposed arrangement goes further in its details, with continuous evaluator access to internal laptops and infrastructure, as well as the ability to publish findings. It is the first part of Amodei’s three-step approach, which also calls for shared standards among AI labs and, later, coordination between governments. Elon Musk responded separately that “Dario is right.”
Safety coordination faces antitrust constraints
Amodei has also called for a narrow US antitrust exemption that would let competing AI companies discuss and coordinate safety standards without that cooperation itself creating competition-law problems.
The legal path is different in the European Union. Brussels has not issued individual exemptions of this kind since 2004, leaving companies to assess their own agreements under Article 101. EU guidance issued in 2023 provides a precedent for competitor cooperation, but its safe-harbor framework concerns sustainability standards rather than AI safety.
The public alignment arrives amid broader concern over increasingly capable systems. An Anthropic researcher recently resigned while warning about a race toward self-improving AI, while OpenAI said at the launch of GPT-6 Astra that future releases could depend more on safety progress than raw performance. The key unresolved issue is whether rival labs can build common safeguards without crossing antitrust lines.
