Anthropic, OpenAI, and Google have been meeting since July about an AI standards body. Nothing is built yet, and small teams should read that carefully.
The Information reported on September 13 that the three biggest frontier labs have held working-group meetings since July 2026 about an industry-led AI safety standards body: shared pre-release testing, independent evaluations, and standardized risk assessments. CNN and the Washington Post followed on September 14. No entity exists, no launch date is set, and the three CEOs still disagree on who should hold the pen. For a five-person shop that buys these models or ships agents on top of them, the useful question is not whether the body gets built. It is who ends up setting your vendors' release bar, and what that does to your calendar and your costs.
The AI story for Monday, September 14, 2026 is a report about meetings, not a meeting.
On Saturday, September 12, Dario Amodei published his essay asking the industry to pace the frontier and to let embedded evaluators verify it. I wrote about that yesterday. It read like an opening move. According to Leo Schwartz at The Information, it was not. Schwartz reported on Sunday, September 13, that Anthropic, OpenAI, and Google have been holding working-group meetings since July to discuss creating an industry-led standards body for AI. The weekend's public statements were built on top of two months of quiet coordination, not the start of it.
CNN's Hadas Gold confirmed the talks with her own sources on Monday morning, and the Washington Post covered them the same day. So the reporting is now three outlets deep. The event itself is older than any of them.
I want to lay out what is actually known, what is still a proposal, and why a small shop that never gets invited to these meetings should still keep a file on them.
What the reporting says, with dates attached
Here is the timeline as best I can reconstruct it from the three pieces. Keep the dates separate, because it is easy to collapse them.
July 14, 2026. Demis Hassabis publishes his proposal for a US-led Frontier AI Standards Body modeled on FINRA: industry-funded, federally overseen, starting with voluntary pre-release reviews that could later become a requirement for the US market. I covered it at the time. CNN's sources say this essay was the catalyst for the talks.
Since July 2026. Representatives from Anthropic, OpenAI, and Google begin meeting regularly in a working group about the standards-body idea. Per The Information's sources, the discussions center on shared protocols for testing frontier models before release: independent evaluations, pre-release safety reviews, and standardized risk assessments. CNN's source says the conversations are ongoing and are continuing with or without the Trump administration.
September 12, 2026. Amodei publishes "We Must Pace the Frontier." Altman and Musk endorse it the same day. The Information's reporting is explicit that the working group predates this.
September 13, 2026. The Information publishes Schwartz's report. This is the first time the extent of the coordination has been reported. Late that Sunday night, Altman posts that OpenAI looks forward to collaborating with colleagues across the industry to formulate the best version of industry standards.
September 14, 2026. CNN and the Washington Post publish their own pieces.
Now the part that matters for anyone tempted to write a headline. As the Pivot News summary puts it, the reporting does not describe a formed entity, a launch date, a governance structure, or which other companies have been approached. None of the three labs has announced anything. Representatives for all three declined to comment or did not respond to CNN. The talks remain talks.
They agree on the diagnosis and disagree on the cure
The three CEOs share one belief: frontier models should face outside scrutiny before public release. They part ways on who should run that scrutiny.
Amodei's model, as Axios laid it out in July, is closer to an FAA for AI. A federal agency with power to block a release. Hassabis wants a FINRA for AI, an industry-funded body under federal oversight that starts voluntary and can harden into mandatory market-access rules. Altman, in a Financial Times op-ed, pushed for something like an IAEA for AI, a US-led international forum that certifies countries, companies, and safety standards.
The Information adds a fourth data point. At an internal OpenAI town hall, Altman reportedly told employees he supports a testing and auditing body for the industry, but believes the major labs will need to build it themselves if the US government does not back it. InvestingLive's summary of the report describes the split this way: Anthropic leans toward partnering with government, OpenAI emphasizes voluntary standards regardless of federal action, and Google DeepMind has been less public about the specifics of its own internal safety protocols.
CNN adds the political layer. Bloomberg reported in July that Treasury Secretary Scott Bessent was considering a FINRA-like independent agency for AI reporting to the SEC. Politico reported earlier this month that Mark Zuckerberg advised President Trump against the idea in a summer phone call. House Speaker Mike Johnson told Jake Tapper on Sunday that there is no consensus among the companies and that Congress is less qualified than the labs to know the ins and outs, so it has to be a partnership.
So: three labs, three institutional models, a Treasury Secretary reportedly sketching a fourth, a Meta CEO against, and a Speaker who says the companies should figure it out with government. That is the actual state of play. Anyone telling you a standards body was launched today is wrong.
Distinct from yesterday, and from the last week
Yesterday's note was about a unilateral commitment. Anthropic said it would give third-party evaluators employee-like access to its own operations, starting now, and asked governments to require the same of others. That is one company changing how it can be inspected.
Today's story is about a multilateral structure that does not exist yet. A standards body would sit between the labs and everyone else, define what "frontier-class" means, run or coordinate pre-release testing, and publish or withhold a judgment. Embedded evaluators are a control inside one lab. A standards body is a shared gate in front of all of them.
They connect, obviously. Amodei's second step in the essay was democratic coordination among labs on common standards, which is roughly what the working group has been discussing. But keep the files separate. The evaluator commitment has a company, a date, and a promised contract. The standards body has a working group and three competing diagrams.
The September 11 note on Altman's antitrust question belongs here too. Competitors meeting since July to agree on shared testing protocols is exactly the kind of coordination OpenAI was asking Congress about. That question is not resolved by this reporting. It is made more concrete by it.
Why a five-person shop should care about meetings it will never attend
You are not in the working group. You will not be consulted on the governance structure. But you are downstream of whatever comes out, in three specific ways.
Who sets your vendors' pre-release bar. Right now each lab runs its own framework. Anthropic has its Responsible Scaling Policy, OpenAI has its Preparedness Framework, Google DeepMind has something less public. When you evaluate a model, you are implicitly trusting that lab's internal bar. A shared body would move that bar outside any single vendor. That is probably good for you. It also means the bar could be set by three companies whose interests are not identical to yours, and whose smallest customers are not in the room.
What it does to your release calendar. Hassabis's original proposal had a 30-day pre-release review window. CNN notes the White House already runs an opaque voluntary version of this. If shared pre-release reviews become normal, model launch dates become review-dependent, and review-dependent dates slip. Do not promise a client a feature tied to a rumored model release. Promise it tied to a model you already have access to.
What it does to your costs and your options. InvestingLive flags the OECD finding that high fixed costs and concentrated infrastructure are already barriers to entry in AI. Standards designed by incumbents tend to fit incumbents. If compliance with a lab-designed body becomes expensive, smaller labs and open-weight providers may fall behind or drop out, and "certified" access tiers may cost more. Your fallback options could narrow just as you need them.
None of that is a reason to panic. It is a reason to write a few things down.
The small-team file
- Sort your vendors by whether they are in the room. Anthropic, OpenAI, and Google are. Meta, xAI, Mistral, and the open-weight providers you use are not, as far as anyone has reported. A standard set by three companies is a standard for their customers first. Note which of your production workflows sit on which side of that line.
- Separate the model's certification from your deployment's. Whatever a future body tests, it will test the model. It will not test your connectors, your permissions, your egress rules, or your human gates. Keep your own evidence for those. This was the point of the July 14 note and it has not changed.
- Put review windows in your planning assumptions. If your roadmap depends on a model version that has not shipped, add a slip buffer. Pre-release review is one more reason a launch date moves.
- Watch the cost pass-through. When a vendor introduces a "certified" or "reviewed" tier, ask what changed in the model versus what changed in the paperwork, and whether the tier you already use is being quietly deprecated.
- Track the promises with dates. Working group: meeting since July 2026, ongoing. Pachocki told reporters earlier this month OpenAI has been talking to external organizations about concrete standards and will have more to share in the next months. Altman on September 12 said more to share soon on evaluators, and late on September 13 said OpenAI looks forward to collaborating on industry standards. Put those in the same file as the evaluator commitments from yesterday, with a review date, and check whether a charter, a name, or a governance structure ever appears.
What not to do
Do not tell your clients the labs announced a regulator. They did not. Do not treat the working group as evidence that standards are imminent; the Frontier Model Forum has existed since 2023 and the Linux Foundation's Agentic AI Foundation since December 2025, and neither one gates a release. Do not confuse the July 14 proposal, the July working-group start, the September 12 essay, and the September 13 report; they are four events, and the working group predates the weekend entirely. And do not skip the vendor sorting because the institutional design debate is more interesting. FAA versus FINRA versus IAEA is a fascinating argument. Which of your workflows run on a lab that is not in the room is a fact you can write down today.
Bottom line
Three frontier labs have been quietly meeting since July about a shared pre-release testing body, the public statements over the weekend were built on that groundwork, and the reporting surfaced on September 13 and 14. Nothing has been formed, and the CEOs still disagree on whether it should be a federal agency, an industry-funded self-regulator, or an international forum. For a Fall River shop shipping client work on these models, the actionable layer is simple: know which of your vendors are in the room, keep your own deployment evidence regardless of what the body eventually certifies, build review slip into your calendar, and track the dated promises until one of them turns into a charter. If you want help sorting your own vendor list this way, that is the kind of work we do.
Sources checked September 14, 2026: Pivot News summary of The Information's September 13 report by Leo Schwartz (September 14, 2026); InvestingLive, Anthropic, OpenAI and Google held talks on AI safety body, report says (September 14, 2026); CNN Business, Top AI companies have discussed creating their own standards body (published September 14, 2026, 7:10 AM ET); Washington Post, Anthropic, OpenAI, Google discussed creating new AI safety body (September 14, 2026); PYMNTS, Google, OpenAI and Anthropic Float Idea of AI Standards Body (September 14, 2026). The Information's original report is gated; claims attributed to it here are drawn from the summaries above.