Anthropic Limits Claude and the API to 185 Countries: What Changes for Teams

If your team builds a product on Claude or lives inside Claude Code, it is time to look at the supported regions page and call a lawyer. Anthropic has moved from a "banned in certain countries" logic to an "allowed only in these 185 regions" logic. The difference is fundamental: anything not on the list is closed by default.

What exactly has changed

The new rules cover Claude, Claude Code and the API. The policy goes beyond a routine check of IP address and location. It has three key points:

— you may not use the products while in an unsupported region, even if you are there on a business trip; — access is closed to legal entities from excluded countries and to their employees, wherever they go online from; — any business that is directly or indirectly controlled from such regions is banned.

The list is based on Anthropic's internal standards and does not replicate US sanctions lists. So the reasoning "there are no sanctions, so it's fine" does not work here. The list includes Iraq, Somalia, Palestine and Taiwan, but not mainland China, Hong Kong or Russia.

Why this is about more than geography

Previously, the typical risk looked simple: a developer from an "inconvenient" country turns on a VPN and hopes nobody notices. Now the risk is structural, and it affects people who don't use any VPN at all.

Imagine a startup registered in a supported country. If it is controlled by a parent company or an owner from an excluded region, then under the wording of the policy such a business is closed too. Another example: your employee flies to a conference or on vacation to a country outside the list and out of habit launches Claude Code in the terminal. That is already a violation, even though the employee and the company are formally in the clear. A third case: a contractor based in an unsupported region gets access to a shared API key or to CI where agents are running.

For product teams, this means the question "where is the server located" is no longer the main one. What matters more is who owns the company and where people access the model from.

What to do in practice

First: take inventory. Find out who uses Claude and from where, where the keys are stored, which CI pipelines and agents call the API, and where contractors physically work.

Second: check your ownership structure against the policy. This is especially relevant for companies with distributed teams and investors from different jurisdictions, since the wording about "indirect control" leaves a lot of room for interpretation.

Third: write rules for business trips and travel. If people travel around the world, they need to know that in some countries their work tool will simply switch off or create legal risks.

Fourth: don't depend on a single provider. Put LLM calls behind an abstraction and keep a tested fallback on another model. A policy-based block can arrive without any technical failure and without warning, and then switching providers over a weekend turns into a scramble.

A caveat: from the public retelling it is unclear how exactly Anthropic will verify compliance with these conditions and what measures it will apply to violators. Read the full original policy text rather than summaries.

What else happened today

— OpenAI launched Control Room: before generating tokens, the inference engine requests permission from it, and a dangerous task will be stopped within 35 seconds. If Control Room is unavailable, all systems stop processing requests. — Google DeepMind's AlphaProof Code formally verified an implementation of the Signal protocol at the level of compiled code.

The analysis took four weeks, humans handled about 1% of the difficult sections, and the system found both of the two deliberately reintroduced fixed bugs. — Nvidia showed the open GR00T 2 model and Isaac Lab 3.0 with the Newton physics engine, as well as the DexBench and VABAR benchmarks for evaluating robot dexterity.

— Google added a Knowledge Work mode to Antigravity, without a terminal or Git, for managers, analysts and designers. The agent retains access to local files and MCP servers.

The key takeaway of the day for practitioners: access to frontier models is becoming a legal issue, not just a technical one. Build this into your architecture and contracts in advance.