What it Means That OpenAI, Anthropic, and Google DeepMind are Coordinating AI Safety — September 16, 2026
· CompaniesAutomation
Radar flash edition: OpenAI confirms via Chris Lehane that it has been coordinating for weeks with Anthropic and Google DeepMind to set AI safety standards, involving independent evaluators and support for the FRONTIER Act. What changes for your company when three providers write the rules for the model you already depend on.
Flash edition. OpenAI has confirmed that it has been coordinating for weeks with Anthropic and Google DeepMind to set AI safety standards among the three of them, without waiting for any government to ask.
What happened
Chris Lehane, OpenAI's head of global affairs, said it on Tuesday, September 15, to the press in Washington: "It is better to try to work together to prioritize safety." Three concrete things are on the table: an industry standards body, the entry of independent evaluators into companies to verify models, and support for the FRONTIER Act, the bill that would force frontier labs to open up to "independent verification organizations" (TechCrunch, September 15, 2026). The thread goes back to July, when Demis Hassabis called for such a body, and Dario Amodei's essay on September 12 calling to slow the pace.
The delicate point is antitrust. Amodei asked for a government exemption; Lehane maintains it is not necessary and cites the precedent of airlines, which share safety information without being a cartel. Not everyone buys it: Aidan Gomez, CEO of Cohere, calls it "a cartel by another name," and FTC Commissioner Andrew Ferguson says he is "deeply leery" (Gizmodo, regarding original reporting from Bloomberg; Reuters wire).
Why it matters
This doesn't stay in Washington. When three providers that concentrate almost the entire market agree on how a model is evaluated, that agreement trickles down to your contract: to the terms of use, to what you can ask an agent, and to what capabilities are withdrawn and when. It goes in the same direction as the AIUC round we covered yesterday, which sells agent auditing like someone selling a SOC 2 (Radar from September 16): third-party verification is already a market. The uncomfortable side is that, if the three agree to slow down, they also slow down your roadmap.
For your company
Two moves, neither expensive. First, inventory your model dependencies: what processes you have in production, with which model, and via which route (direct API, Azure, an intermediary, or a tool where you don't even know what it uses internally). If a revenue-generating process depends on a single model from a single provider, that's a continuity risk, not just a price risk. Second, run the same prompt through two models from different houses and save the result: not to switch, but to know how much it would cost you to switch. It's the cheapest insurance policy against a third-party policy change. The lock on the AI First ladder: this doesn't move you up a rung, it protects the one you already have. Without any processes in production, there's still nothing to do here.
Frequently Asked Questions
Does this affect a small business using ChatGPT or Claude?
Not directly today. But the standards that emerge from there end up converted into contract clauses and capabilities that are activated or withdrawn. If you have a serious process built on a model, you should care who is writing those rules.
Is it legal for three competitors to coordinate like this?
That is the ongoing debate. OpenAI says it doesn't need an antitrust exemption; Anthropic preferred an explicit one; the FTC and Cohere see a risk of leaders writing the rules to suit themselves. Without an approved law, everything is voluntary.
What is the FRONTIER Act?
A U.S. bill that would require frontier model labs to undergo safety audits by independent verification organizations. it is not passed and its timeline is uncertain.