Three Rival Labs Explore a Joint Frontier AI Standards Body

Anthropic, OpenAI and Google have been holding working-group discussions since July about forming a joint industry body to set safety standards for frontier AI models, according to the proposals now circulating among the three labs. The talks predate the public call by Anthropic chief executive Dario Amodei for AI companies to slow the pace of capability development and coordinate on safety. At the centre of the debate sits a proposal from Google DeepMind leadership for a US-based frontier AI standards body modelled on the Financial Industry Regulatory Authority, the self-regulatory organisation that oversees US broker-dealers.

The plan would hand the largest frontier laboratories a formal role in writing the rules their rivals must clear before launch. That is a striking shift for three companies that compete directly on model capability, and it frames the question that will decide whether the effort survives: can a body built by the incumbents work as neutral infrastructure, or does it become a gatekeeping mechanism that locks in their lead?

What a Frontier AI Standards Body Would Actually Do

The most concrete element under discussion is a pre-release assessment window. Labs would submit frontier models for technical testing and auditing up to 30 days before release, with the results feeding a shared evaluation standard that every participant recognises.

The governance design borrows from FINRA. Under the DeepMind proposal, the frontier AI standards body would be US-initiated and funded mostly by industry, with independent technical experts and open-source representatives seated on its board. Anthropic has floated a second mechanism: an internal but independent ombudsman inside each AI company, equipped with checkpoints to verify that developers follow standardised guidelines and insulated from commercial pressure.

Neither the pass thresholds nor the consequence for a lab that fails an assessment has been settled. Without those two pieces, the standard describes a process rather than a rule.

A Crowded Field of Existing Bodies

A new body would land in territory that is already occupied. Five separate efforts now touch frontier model governance, which raises the question of whether a sixth adds enforcement capacity or just another set of meetings.

The pattern across those initiatives is voluntary participation and shared technical vocabulary. None can compel a lab to delay a launch, which is the gap the FINRA template is designed to close. Building another body on top also carries a coordination cost: the same small pool of safety researchers and standards staff gets spread across more boards and working groups.

Where the Three Labs Diverge

Agreement runs to diagnosis, not to machinery. All three companies accept independent pre-release testing and a single standards body. They split on how much formal authority governments should hold. Anthropic favours a partnership with regulators, while OpenAI leans toward voluntary industry rules with lighter state involvement. That fault line mirrors the split visible in US state-level AI safety debates, where mandatory third-party audits and industry self-attestation have competed as rival models.

Google DeepMind's FINRA template is a partial compromise. FINRA is technically a private organisation, yet it operates under statutory authority delegated by the Securities and Exchange Commission, and its rules carry enforceable weight. Importing that structure into AI would require Congress to delegate comparable authority, a far heavier lift than the voluntary framing OpenAI prefers. Absent that delegation, a FINRA-style frontier AI standards body would take the shape of a regulator without the teeth.

The ombudsman idea and the pre-release window solve different problems. An internal monitor sees training runs as they happen, months before a model is finished, but holds no power to stop a release. A pre-release audit arrives late enough to catch failures and carries no remedy unless the body can impose one. The two mechanisms together would cover more of the pipeline than either alone, which is why both remain on the table even though they come from different companies.

How Fast the Clock Is Running

Anthropic's chief executive has put a short timeline on the problem, warning that within six to 12 months a coordinated swarm of agents could reach a scale where it is capable of taking over the entire internet. That warning is the strongest argument the three labs have for moving quickly, and it is also the reason a purely voluntary standard may not hold.

Sam Altman of OpenAI and Demis Hassabis of Google DeepMind have each called for a new international body to vet and set standards for cutting-edge AI. The shared position gives the talks political cover in Washington, where the alternative has been a patchwork of state rules rather than a federal framework.

Two Tracks, One Compliance Load

The EU's voluntary General-Purpose AI Code of Practice already covers Anthropic, Google and OpenAI. A US frontier AI standards body would sit alongside it rather than replace it, and that means labs shipping globally could face two sets of definitions, two testing protocols and two documentation regimes.

The European track also weakens the case for pure self-regulation. All three companies signed a written code when regulators applied pressure, which shows the industry will accept an external framework when the alternative is statutory law. OpenAI's preference for voluntary rules is therefore a position about who writes the standard, not about whether one should exist.

The Compliance-Cost Problem

The sharpest objection to common standards is economic. A mandatory 30-day assessment window, plus the documentation and testing a shared standard would require, carries a fixed cost that scales badly for smaller labs.

For Anthropic, OpenAI and Google, absorbing that cost is routine. For a startup working through a single frontier training run on limited compute, a month-long hold before release can consume a meaningful share of its runway and hand a timing advantage to whoever is already ahead. The same three companies drafting the standard also hold the largest evaluation teams and the deepest bench of safety researchers, which makes it easier for them to clear a bar they helped write.

That dynamic is not unique to AI. Financial regulation built around broker-dealers has long drawn criticism for favouring large incumbents that can spread compliance costs across more revenue. Importing the FINRA model imports that criticism with it.

A Bloc That Is Not United

The three conveners do not agree on model access either. Twenty-five technology companies have backed an open-weight AI letter, and Anthropic is not among the signatories, while Google ships open-weight Gemma models that have passed 300 million downloads. A shared standard would have to say something about open-weight releases, and that is where the coalition is most likely to fracture.

Open-weight releases can be fine-tuned and redistributed without the original lab's involvement, so a testing regime that stops at the developer's door leaves the hardest cases outside its reach.

Why this matters

If a shared frontier AI standards body emerges with real pre-release testing, it becomes the de facto gate for anyone shipping a frontier model, and the three labs that convened the talks would sit closest to the gate. For enterprise buyers and investors, the practical signal is timing: a 30-day assessment window reshapes release schedules, procurement planning and how fast new capabilities reach production. For smaller developers, the question is whether they get a seat in the working groups or inherit a standard written without them.

Photo by Brecht Corbeel on Unsplash

Related Articles

• Three Labs, One Opt-Out: Inside the White House AI Review Framework
• Washington's Frontier AI Regulation Plan Borrows from FINRA
• AI Model Fatigue Grips Buyers as Four Frontier Labs Ship in One Week

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.