Anthropic, OpenAI, and Google Propose Private AI Standards Body
Workgroups from the three companies have met since July to establish shared safety benchmarks, transparency protocols, and external auditing.
Three labs that have been publicly feuding for the past two years have been sitting at the same table since July. Workgroups from Anthropic, OpenAI, and Google are discussing the creation of a voluntary standards body for frontier models, without direct involvement from the U.S. federal government. This information, first reported by The Information and echoed by Asian media on September 14, formalizes what Sam Altman hinted to Fortune 48 hours earlier: "some agreement among major AI companies is going to happen."
What the body would and wouldn’t do
According to sources connected to the three companies, the proposed structure includes shared safety benchmarks, cross-red teaming protocols, transparency clauses before launching models above a certain size, and an external auditing mechanism similar to what Dario Amodei proposed to CBS on Sunday. What is not included: any power of sanction. It is a club of commitment, not a regulator. It does not replace the European AI Act or the Frontier Model Forum, which remains on paper.
The direct trigger was the stalemate in talks with the White House. The workgroups began in July precisely because meetings with David Sacks’s team, the Trump administration's AI czar, stopped producing results. Sacks has publicly argued that any deceleration agreement among American labs is "unilateral disarmament" in the face of China. The companies concluded that waiting for Washington is betting against their own launch schedules.
Where the three disagree
Anthropic and OpenAI disagree on the most obvious point: the level of state involvement. Anthropic has argued since July, through its head of red team, that standards need to be co-developed with regulators. OpenAI advocates the opposite: voluntary industry standards alongside targeted state legislation (California's SB 1047 is the model). Google occupies a middle ground, operationally closer to OpenAI while publicly aligning with Anthropic.
The distance matters because it defines whether the body could become a vehicle for regulatory capture. A fully voluntary club establishes a baseline that only those with revenue to support expensive red teams can meet, benefiting the three founders. A club that requires government committee involvement creates a formal interlocutor with whom any future government (or Congress itself) can engage, which benefits those who fear political revision of the arrangement in the future.
Global implications
For the United States, the structure solidifies a trend that Blockchain Association and IBM experienced in previous cycles: industrial self-regulation precedes regulation and often defines it. If the body is formed, the Frontier Model Forum becomes redundant, and the Department of Commerce loses the central reason for its current execution of export controls against China.
For the European Union, the AI Office gains a unique interlocutor to hold accountable. This is operationally good and politically bad: Europeans wanted the AI Act to be the global baseline, and an American club with its own standards creates a second measure. French Minister Clara Chappaz already indicated in July that Paris would treat "American voluntary standards" as competition, not convergence.
For Brazil, the effect is indirect and costly. Bill 2338/2023, approved in the Senate in December 2024 and still stalled in the Chamber, was designed to align with the European AI Act. If the standard changes, the Brazilian text finds itself in limbo: neither risk-based in the European manner nor voluntary in the American way. Multinationals operating here, notably Itaú, Bradesco, Petrobras, and Ambev, will demand contractual clauses compatible with the American club from their AI suppliers. Anthropic and OpenAI are already charging a premium for enterprise versions with "safety attestation"; this premium is likely to rise.
What could still derail the agreement
The first risk is Sacks himself. If Washington signals that any coordination among labs could become an antitrust case (the Department of Justice is already monitoring pricing discussions in image models), the companies will back off. The second risk is Elon Musk’s xAI, which is not involved in the talks and may personally incentivize accusations of collusion against the trio. The third, more gradual risk: if the next frontier model is launched before the body publishes its first standard, the standard becomes useless. The workgroups are discussing a public version in January. GPT-6, Fable 5.2, and Gemini 4 are informally slated for release before that.
For the CIO acquiring a model today, the practical takeaway precedes any club: write into the contract the right to audit the red team, demand SLA for disclosure within 72 hours for safety incidents, and treat the "safety attestation" as a clause, not as a sales argument.