Anthropic released a "Support Region Policy" that changed the product's availability from "where it cannot be used" to "only where it can be used on the list." There are 185 countries and territories on the list, which takes effect on November 12.
The products covered are comprehensive: Claude.ai, Claude Code, Cowork, API, and Claude models provided through cloud platforms and authorized resellers, all follow this list.
In addition to the list itself, there are three prohibitive provisions in the policy, which are written in more detail than before:
It is illegal to use it when you are in an unsupported area. Supporting employees of regional companies to travel to these places also counts.
It is illegal to use it by entities registered or headquartered in unsupported areas and their personnel, regardless of where the person is.
The personnel of companies that are directly or indirectly majority-owned or controlled by entities from non-supporting areas are also in violation.
The list includes Taiwan, China. Ukraine excluded the five regions of Crimea, Donetsk, Kherson, Luhansk and Zaporozhye. Mainland China, Hong Kong, and Macau are not included.
Another detail is that Iraq, Libya, Somalia, Sudan, and Palestine are on the list. So it does not copy the US sanctions list, but Anthropic’s own set of compliance standards. Dependent territories and dependencies are treated as the mother country. For example, Puerto Rico is treated as the United States.
For domestic developers, the actual impact is mainly on Articles 2 and 3. The previous restrictions looked at the source of access, but now the company's registration place and shareholding structure are clearly included. If the main body of the overseas team is in China, or is controlled by a domestic entity, even if the employees and servers are overseas, it will still fall within the scope according to the provisions.
Teams that rely on Claude API to build products should prepare a switchable model solution now. There are already many choices in domestic models that have decent code and long text capabilities. The interface layer is well abstracted, so changing models will not be a hassle.