• OpenAI and Anthropic have reportedly moved toward an agreement to stress-test each other's frontier AI systems, according to people familiar with the matter.
  • The talks, which also include Google DeepMind (GOOG), signal a shift from self-assessment to rival-led evaluation in the race to deploy safer AI.
  • While no binding deal has been finalized, the discussions could set a precedent for independent oversight in the rapidly evolving AI industry.

Rivals Team Up on Safety

In an unusual twist for the fiercely competitive world of artificial intelligence, OpenAI and Anthropic are reportedly nearing an arrangement to stress-test each other's AI models—a move that would mark a significant step toward independent evaluation of high-risk systems. The discussions, first reported by Bloomberg, have been underway for several weeks and also involve Google DeepMind, according to Chris Lehane, OpenAI's global policy chief. Lehane told Reuters on September 15 that OpenAI had been in talks with its rivals about AI safety but dismissed the need for an antitrust waiver to coordinate on such efforts.

The potential cross-testing deal comes as pressure mounts on AI developers to prove that their systems are safe before deployment. Rather than relying on internal assessments, the arrangement would allow a direct competitor to probe for dangerous capabilities, security vulnerabilities, or misuse risks. "What institutional investors like us are really focused on is regulatory stability," said Andrea Valeri, Blackstone (BX)'s country chairman for Italy, at a recent conference—a sentiment that echoes across the tech industry, where regulatory uncertainty looms large.

A Shift Toward External Scrutiny

The reported talks reflect a broader movement from voluntary safety pledges to more operational mechanisms, such as embedded third-party evaluators and pre-deployment testing. Anthropic CEO Dario Amodei has been a vocal advocate for pacing capability advances and has committed his company to permanent third-party evaluators with employee-level access. OpenAI's Sam Altman has signaled openness to a similar commitment.

"We have a constant balance with the banks, which really we consider our partners and not only our binary competitors," said Cecile Mayer-Levi, head of private debt at Tikehau Capital (TKKHF), speaking about a different sector but capturing the spirit of cooperation that could apply to AI safety. "It's much more of a convergence between the two solutions."

Elon Musk has separately proposed that major AI developers, including OpenAI, Anthropic, Google, Meta (META), xAI, and leading Chinese firms, run each other's test harnesses. However, no named rival had publicly agreed to that wider proposal at the time of reporting.

Financial and Strategic Stakes

Neither OpenAI nor Anthropic is publicly traded, but both are preparing for potentially massive IPOs, according to Reuters. That raises the commercial stakes around demonstrating that their systems are both capable and responsibly governed. The tension is palpable: frontier-model development demands enormous expenditures on chips, data centers, energy, and talent, yet better safety processes could raise short-term costs and slow releases.

The arrangement, if finalized, could also have antitrust implications. Amodei has argued that certain collective steps might require targeted U.S. antitrust exemptions, while Lehane has said the current discussions do not require one. The Trump administration has resisted broad new AI regulations, framing rapid development as key to U.S. competitiveness against China. National Economic Council Director Kevin Hassett said the private sector should address many concerns, with government monitoring and law enforcement where necessary.

Regulatory and Geopolitical Hurdles

The path to a binding agreement is fraught with challenges. A Chinese Foreign Ministry spokesperson characterized calls for an AI slowdown as "fear mongering," underscoring the difficulty of building a global safety regime while AI capability is viewed as a strategic asset. Amodei has advocated for tighter controls over advanced AI chips and model weights, but international cooperation remains elusive.

In the near term, talks among OpenAI, Anthropic, and Google DeepMind are expected to continue, with pressure to clarify whether they will produce a formal protocol or merely nonbinding dialogue. The most plausible early implementation would involve limited pre-release access under strict confidentiality controls—not unrestricted sharing of model weights or proprietary data. Politically, scrutiny will grow around whether voluntary collaboration is sufficient and whether disclosures from testing should be public or confined to regulators.

A spokesperson for Anthropic declined to comment on the specifics of the talks. OpenAI did not respond to a request for comment by press time.

Correction: An earlier version of this article misstated the timing of Chris Lehane's remarks. He spoke on September 15, not September 5.