Anthropic and OpenAI rivalry fades as industry shifts focus to broader AI landscape

The U.S. government is preparing to exert significant control over the release of AI models.

Two weeks after the U.S. government halted Anthropic’s Fable and Mythos models, OpenAI’s latest release appears to face similar restrictions. The Information reported Thursday that GPT 5.6 will be launched in a limited preview, with government approval granted “customer by customer” until a broader release is authorized.

If this preview lasts only a “couple of weeks,” as Altman reportedly suggested, it may not pose a major issue. However, Mythos has already been in preview for months, with no clear timeline for general release. Even a few weeks of review could significantly reduce the economic benefits of a costly new system, especially as AI labs struggle to improve their financial performance. If model development slows, it could also dampen ongoing data center expansion.

If this situation worsens, the entire industry could be at risk.

Critically, OpenAI and Anthropic now face identical challenges and potential consequences. Industry discussions often focus on blaming one side, with accusations ranging from Anthropic’s alleged regulatory capture to OpenAI’s close ties with Trump to eliminate competition. This is understandable, as many industry leaders have billions of dollars at stake.

However, the current situation extends beyond individual company disputes. The cost of implementing a disjointed government approval process for every frontier model is evident, and there is no solution that benefits one lab without helping others.

The immediate challenge is establishing a sensible release process. While government testing before release is standard for many consumer products, GMU fellow (and soon-to-be OpenAI employee) Dean Ball highlighted in a detailed post this morning that it remains unclear what safety assurances would satisfy regulators. The U.S. government lacks the expertise and capacity for the necessary testing. Moreover, there is no clear articulation of what risks regulators are trying to mitigate.

It is tempting to view the government process as the sole problem, but underlying concerns are valid. Even if you dismiss the Mythos hype, there is clear evidence that AI tools are transforming cybersecurity. Similar dynamics apply to biorisk and alignment. Restricting model releases alone is not the solution, as it limits public access, but genuine concerns must be addressed.

The best solutions, as outlined by Ball, require collaboration. This means trusting independent groups to guide the process, even if their goals do not fully align with yours. It involves supporting the least problematic regulatory options available, rather than opposing every regulation. Most importantly, it means advocating for AI as an industry, rather than treating safety and regulation as opportunities for competitive advantage.

For many in AI, this will be a difficult pitch. Unfortunately, AI models have advanced to the point where their capabilities have real political implications. Addressing these consequences requires collective action. In the coming weeks, we will see if the industry can achieve this.