OpenAI, Anthropic and Google DeepMind confirmed on September 15, 2026 that they had spent several weeks collaborating on a shared safety framework for artificial intelligence (AI). This rare alliance between longtime rivals underscores growing concerns over catastrophic risks from frontier models, while raising significant antitrust questions and drawing sharp criticism from smaller industry players.
🔑 Key takeaways
- OpenAI, Anthropic and Google DeepMind collaborated for several weeks on AI safety
- The initiative follows an essay by Anthropic CEO Dario Amodei urging the industry to slow frontier development
- US senators are proposing an antitrust exemption to legalize the cooperation
- The three giants also aim to counter distillation attacks by Chinese laboratories
- Smaller players like Cohere denounce a « cartel in disguise » fearing market distortion
An unprecedented alliance between historic rivals
The announcement was made by Chris Lehane, OpenAI’s head of global policy, during a press briefing in Washington. « It is better to try to work together to prioritize safety, » he said, according to reporting by Bloomberg and Reuters. The cooperation follows an essay published on September 12, 2026 by Dario Amodei, CEO of Anthropic, in which he called on the industry to « slow the rapid pace of frontier model development » to avoid « catastrophic risks. »
Support was almost immediate. Sam Altman (OpenAI), Elon Musk (SpaceXAI Corp.) and Demis Hassabis (Google DeepMind) all backed Amodei’s proposal. OpenAI notably said it was ready to embed third-party evaluators within its development teams to monitor model safety. This mobilization stands in stark contrast with the fierce commercial battle the three players usually wage in the generative AI market.
The antitrust puzzle
The cooperation between the three giants raises major legal questions. Several experts warned that the coordination could be classified as an « output restriction » under the Sherman Act, which would constitute a violation of US antitrust law. To defuse this legal time bomb, senators Jim Banks (Indiana) and Adam Schiff (California) introduced the « Collaboration on Adversarial Threats and Security Risks Act, » a bill that would explicitly authorize AI labs to cooperate on safety.
« Once again, under the pretext of protecting the public, these oligopolies are asking to bend competition rules and dictate terms for everyone else. A wolf in sheep’s clothing, a cartel by another name. »
Aidan Gomez, CEO of Cohere
Dario Amodei himself suggested that the US government grant an antitrust exemption to facilitate cooperation. Chris Lehane however argued on September 15 that such a measure was unnecessary, while reiterating OpenAI’s support for two bipartisan bills: the FRONTIER Act, which requires independent audits before new models reach the market, and a second bill led by Senate Majority Leader John Thune alongside Ted Cruz (R) and Amy Klobuchar (D), empowering the Secretary of Commerce to order such audits.
A common front against Chinese distillation
Beyond safety, the cooperation between OpenAI, Anthropic and Google targets another strategic objective: countering distillation attacks carried out by Chinese labs. The technique consists of mass-querying a high-performing US model, collecting its responses, and reusing them to train a cheaper competing model. According to Tech Brew, Microsoft already raised the alarm in January 2025 by accusing Chinese startup DeepSeek of extracting large amounts of data through OpenAI’s API.
Anthropic for its part identified more than 24,000 fake accounts used by three Chinese companies on Claude, generating 16 million exchanges, some traced back to senior executives at those firms. US officials estimate the cost of these practices at « billions of dollars » per year. To share intelligence, the three labs will use the Frontier Model Forum, a non-profit they co-founded with Microsoft in 2023. In parallel, Apple currently pays Google roughly $1 billion per year to use Gemini in an upgraded version of Siri, illustrating the growing economic weight of AI alliances.
Industry positions at a glance
The initiative is far from consensual. On Truth Social, US President Donald Trump rejected Amodei’s call, arguing that any slowdown in AI development would let China take the lead in the tech race. His AI advisor David Sacks dismissed existential concerns as « exaggerated. » The release of DeepSeek’s reasoning model in January 2025, matching US models on several benchmarks, had already wiped nearly $1 trillion off tech valuations in a single day.
| Player | Position | Main argument |
|---|---|---|
| OpenAI | In favor | Common framework and third-party evaluators are essential |
| Anthropic | In favor | Catastrophic risks, calls for antitrust exemption |
| Google DeepMind | In favor | Public backing, coordination through the Frontier Model Forum |
| Cohere | Opposed | Cartel in disguise, competition distortion |
| Trump administration | Opposed | Risk of losing leadership against China |
| Senators Banks & Schiff | In favor (legislative) | Legalize cooperation through a bipartisan bill |
Conclusion: lasting framework or temporary alliance?
The cooperation between OpenAI, Anthropic and Google DeepMind marks a turning point in frontier AI governance. By joining forces on safety and against distillation, the three labs effectively acknowledge that existential risks and geopolitical competition outweigh individual commercial interests. Whether the alliance survives headwinds remains to be seen: antitrust pressure, political opposition, and distrust from smaller players like Cohere all stand in the way.
The coming weeks will be decisive. The fate of the bipartisan bills, the evolution of distillation attacks, and the final US government stance will determine whether this pact turns into a durable regulatory framework or remains a coalition of convenience. One thing is certain: the frontier AI landscape has entered a new phase of maturity in which safety is no longer a marketing argument, but a shared strategic stake on a global scale.
Sources
This article is published for informational and educational purposes only. It does not constitute investment advice. Do your own research (DYOR) before making any decision.
