The U.S. government has again allowed some companies to use Anthropic's latest model, "Mythos 5." It comes about two weeks after U.S. Commerce Secretary Howard Lutnick sent a letter on June 12 (local time) demanding that foreign nationals be denied access. Whether the same company's general-purpose model "Fable 5" will be granted similar relief had not been decided as of the 26th.
The June 12 measure was broad in scope. It blocked access by overseas users as well as by foreign nationals staying in the United States. Barring people of a particular nationality from using a U.S. company's service inside the United States differs in character from existing export controls. The object of control has moved from software to users.
The repercussions spread beyond the border. The question of access to Mythos was taken up as an agenda item at the G7 summit held in mid-June. Meanwhile, OpenAI announced on June 22 that it was expanding partnerships to Australia, Canada, France, Germany, Japan, Poland and South Korea, along with EU organizations.
Performance figures were released as well. The updated version of "GPT-5.5-Cyber," which OpenAI released on June 22, scored 85.6% on the company's internal evaluation tool "CyberGym." CyberGym is a testbed that measures how well a model finds security vulnerabilities. Mythos 5 scored 83.8% on the same evaluation.
A regulatory design was also put forward. Google issued a white paper on June 25 titled "A Practical Approach to America's AI Governance," proposing that the most advanced frontier models and the applied services spread through daily life be treated separately. For frontier models, it calls for establishing an independent safety body under federal oversight to verify companies' own safety audits. The design is modeled on the oversight approaches used in nuclear power, finance and aviation.
For consumer services, it chose revising existing law over new comprehensive regulation. That means modernizing laws on child protection, copyright, the labor market and personal data. It also included items calling for limits on features in which chatbots pose as humans or induce emotional dependence, a ban on repeated rewards and gamification elements in services for minors, and a requirement to provide counseling agency information in response to questions about suicide and self-harm. On the power grid and data centers, it held that a response on the scale of the Eisenhower highway program is needed.
The regulatory debate also moved into elections. In the June 23 Democratic primary for the U.S. House in New York's 12th District, state Assembly member Alex Bores, who had co-sponsored a bill requiring safety measures at AI companies, took 35.0% and lost to candidate Micah Lasher, who took 39.1%. A $100 million super PAC backed by executives from OpenAI, Palantir and Andreessen Horowitz spent $8.15 million to defeat Bores.
On the other side, the safety and regulation camp, including Anthropic, put in $19.26 million. That amounts to $27.41 million concentrated in a single district.
China moved along a different path. The State Administration for Market Regulation announced a national standard for "AI agent interconnection" on June 26. It consists of seven detailed standards covering the overall architecture and the assignment of unique identity codes, with the aim of a closed system binding all AI agents into a single identity management framework. An AI agent refers to software that takes instructions from a person and carries out tasks on its own.
A similar design has been applied to robots as well. The standardization committee under the Ministry of Industry and Information Technology launched a "humanoid full-lifecycle management service platform" in May and is assigning unchangeable unique identification numbers to bipedal AI robots. Whether the assignment of these identification numbers will be mandatory, and when it would take effect, has not been announced.
The U.S. AI action plan released in July 2025 set out accelerated innovation, infrastructure and the international spread of the technology. China has pushed the integration of manufacturing and health care and the spread of open source through its 2017 development plan and the 2025 "Artificial Intelligence Plus" initiative. Meanwhile, the money U.S. big tech is putting into model development and data centers this year exceeds $700 billion. Neither side treats regulation solely as a safety matter.
The disposition of Fable 5 is a matter to be decided soon. The outcome will determine which countries' researchers can use which models, and which companies' chatbots can enter which markets.
