OpenAI, Anthropic and Google DeepMind have been in discussions on AI safety for several weeks, according to OpenAI global policy chief Chris Lehane. The talks arrive as the largest frontier-model developers face mounting pressure to show that safety commitments can be translated into repeatable, externally credible controls.
The immediate news is coordination among direct competitors. The more consequential question is whether that coordination produces shared evaluation methods, disclosure practices or an industry standards body—and whether those mechanisms can withstand legal and political scrutiny.
What changed
Lehane said OpenAI has been working with Anthropic and Google DeepMind on safety issues, while in Washington to engage lawmakers on catastrophic AI risks. The disclosure follows Anthropic CEO Dario Amodei’s call for frontier AI companies to cooperate on slowing development when risks warrant it.
OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis have publicly supported the broader call for action. Altman also said OpenAI would join Anthropic in using third-party evaluators to monitor safety.
Separately, reporting has indicated that the three companies have discussed creating an AI standards body. Hassabis has previously advocated for a U.S. watchdog able to assess the most advanced models and coordinate industry-wide slowdowns if risks escalate.
Why operators should care
For companies building on frontier models, safety governance is increasingly likely to become part of procurement, deployment and vendor-risk work—not just a policy debate.
If leading labs converge on common model evaluations or independent assessment practices, enterprise buyers may gain more comparable evidence about a model’s capabilities and failure modes. That could eventually shape requirements for high-impact uses such as code generation, cybersecurity workflows, autonomous agents and systems handling sensitive business data.
It also changes the planning calculus for builders. A coordinated safety framework could introduce more predictable release gates, but it could also affect model availability, feature rollouts and the documentation customers receive. Teams should avoid assuming that access to increasingly capable models will be uninterrupted or that safety practices will remain unique to each vendor.
For founders, a common standard could reduce the cost of responding to different vendor and customer expectations. But a standards process dominated by the largest labs may also set compliance norms that are difficult for smaller model developers and application companies to meet.
The antitrust boundary matters
Competitor coordination is not automatically unlawful, but its scope is crucial. Altman has acknowledged that joint safety discussions could raise antitrust concerns if they are viewed as suppressing competition. Amodei proposed a narrow government waiver for safety coordination; Lehane reportedly said such a waiver is not needed.
That distinction will determine whether the effort can move beyond broad principles. Sharing technical evidence on severe risks or aligning on independent testing may be easier to defend than agreements that effectively determine when products can ship or which capabilities can be offered.
The policy environment is another constraint. OpenAI has backed a provision in the proposed FRONTIER Act that would require leading labs to permit independent verification organizations to assess safety practices. But the Trump administration has pushed back against arguments for tighter AI restrictions, framing slowdowns as a competitive disadvantage relative to China.
What to watch next
The useful signals will be operational rather than rhetorical:
- **A defined structure:** whether the companies announce a standards body, its membership, governance and funding.
- **Shared evaluation protocols:** whether they publish common thresholds, test categories or reporting formats.
- **Independent access:** whether third-party evaluators receive meaningful access to models, systems and evidence rather than conducting limited reviews.
- **Legal guardrails:** whether participants publish antitrust protocols or seek clearer guidance from regulators.
- **Customer-facing commitments:** whether safety discussions yield release notes, incident processes, audit materials or contractual assurances that enterprise teams can use.
For now, the talks demonstrate that frontier AI safety is becoming a competitive and operational issue at the same time. The test is whether the companies can create credible safeguards without turning industry coordination into a mechanism that regulators—or customers—see as closed, self-policing control.
