OpenAI has been working with Anthropic and Google DeepMind on AI safety issues, according to reporting on September 15, 2026. OpenAI global policy chief Chris Lehane said the discussions had been underway for several weeks. The important point is what has actually been confirmed: cooperation on safety discussions, not a finished regulator, shared model development program or agreement to stop AI development.
What has been reported
Reuters, citing Bloomberg, reported that OpenAI is working with Anthropic and Google DeepMind on AI safety. TechCrunch separately reported that Lehane described the talks as having continued for weeks. Reuters said it could not independently verify the report and noted that the companies had not immediately responded to its requests for comment.
That attribution matters. The public evidence supports saying OpenAI's policy chief described ongoing discussions. It does not support turning every reported proposal into a completed industry agreement.
How the talks connect to the wider safety debate
The discussions followed a public essay by Anthropic CEO Dario Amodei calling for the industry to slow the pace of frontier AI development so safety measures can catch up. AP reported that Amodei proposed stronger external evaluation and international coordination. OpenAI CEO Sam Altman has expressed support for safety measures, while the broader industry response is not uniform.
Why external evaluation keeps appearing
The reported proposals include independent or third-party evaluation of advanced AI systems. The logic is straightforward: a company testing its own model has access to technical information but also has a commercial interest in shipping products. External evaluation could add another layer of scrutiny, although the practical design—what evaluators can access, who pays, what gets published and how security-sensitive information is protected—remains difficult.
Why an antitrust question is part of the story
Lehane said OpenAI did not believe the three companies needed an antitrust waiver to coordinate on safety matters. This is narrower than saying all cooperation among competitors is legally approved. The exact scope of lawful cooperation depends on what companies share and how the collaboration is structured.
Meta's position shows the industry is not aligned
AP reported that Meta CEO Mark Zuckerberg distanced Meta from calls for a coordinated slowdown. Zuckerberg argued that companies should remain individually responsible for safety and pointed to Meta's decision to delay its Muse AI agent for additional safety work.
This contrast is useful context. The industry discussion is not a simple agreement among AI labs. Different companies can support evaluations or safety work while disagreeing about a coordinated pause, regulation or the pace of development.
What this means for AI product teams
For developers and businesses, the practical signal is that safety controls are moving closer to deployment infrastructure. Model evaluations, access controls, monitoring, approval systems and incident response are becoming part of the product stack. But the public reporting does not establish a single shared standard that companies can adopt immediately.
What remains unknown
There is no public technical specification for the reported OpenAI-Anthropic-Google safety collaboration in the sources reviewed here. The exact governance structure, evaluator access, testing methods, shared datasets and final outputs are not established. Those unknowns should remain separate from the confirmed fact that discussions have taken place.
The evidence-based takeaway
The most meaningful development is not that competitors suddenly agree on AI policy. It is that major labs are discussing shared safety work while disagreeing on how much coordination and slowdown is appropriate. That makes evaluation, governance and accountability important areas to watch as frontier models become more capable.