OpenAI, Anthropic and DeepMind held safety talks without Meta

What OpenAI confirmed and what it did not

OpenAI went on record on 15 September 2026 confirming it has been in talks with Anthropic and Google DeepMind on AI safety for several weeks.^[1] It is the first time any of the three labs has publicly acknowledged the discussions are happening.

OpenAI did not confirm any formal structure, any agreed timeline, or whether the talks will produce anything binding. The scope stays vague. OpenAI said it does not see the need for an antitrust waiver to coordinate safety efforts with Anthropic and Google DeepMind.^[1]

OpenAI also said it would support bipartisan legislation targeting catastrophic AI risks, though no such legislation currently exists.^[1]

What DeepMind proposed two months earlier

The talks did not start in September. The groundwork was laid on 14 July 2026, when Google DeepMind CEO Demis Hassabis published a detailed proposal on X outlining a new oversight structure for frontier AI.

Hassabis proposed a FINRA-style independent standards body that would require labs to submit models 30 days before release for independent testing, update safety benchmarks quarterly, and hold the power to coordinate a slowdown if risks escalate.^[2] The FINRA comparison is deliberate: the US Financial Industry Regulatory Authority sits outside government but carries genuine authority over the firms it oversees. Hassabis is pitching something with teeth, not a voluntary code of conduct.

Anthropic's position in the talks has not been detailed publicly. Anthropic has neither confirmed nor denied what it is willing to accept under any proposed framework.

Meta's absence and Zuckerberg's said position

On the same day OpenAI confirmed the talks, Mark Zuckerberg wrote on X that the industry does not need coordinated oversight at all.

Zuckerberg said that every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens.^[3] The argument is a market-incentives case: legal liability and competitive pressure already motivate responsible development, so a shared oversight body adds cost without adding safety.

Meta has the most to gain from moving fast. Its Llama model family is open-source, and any coordinated slowdown or pre-release testing regime would hit Meta's release cadence harder than it would hit closed-model competitors. The incentive structure here is not subtle.

Why self-regulation ahead of legislation matters

No formal AI safety legislation is in place, making this voluntary coordination the only safety net currently on the table for frontier model releases.^[1] That absence of law is exactly why the labs have room to write their own rules. Governments move slowly. Labs release models every few months.

A voluntary framework that excludes the world's largest open-source AI publisher covers a shrinking share of the models actually being deployed. Even if OpenAI, Anthropic and DeepMind agree on pre-release testing, Meta can ship whatever it builds on its own schedule. Ask 'and then what?' and the answer is: the three labs have a code of practice and Meta has a competitive advantage.

The Hassabis proposal gives the clearest picture of what a real framework could look like: 30-day pre-release windows, quarterly benchmark updates, a coordinated slowdown mechanism.^[2] Whether OpenAI and Anthropic are willing to accept those specific terms is still unknown. The talks that started several weeks ago have not produced a formal structure, a name for any standards body, or a date by which one might be announced.^[1]