U.S. designates Anthropic AI for export controls, halts access

2026-06-24

Anthropic's latest AI, Claude Mythos 5, was designated for U.S. export controls after showing strong performance in advanced tasks close to cyberattacks. In response, the company halted access, including for U.S. users. Attention on Japanese companies and domestic AI is also growing.

Mythos 5 performance

Mythos 5 was announced on U.S. time on the 9th. Anthropic stopped offering it the same day after the U.S. government designated it for export controls on the 12th. Developers often publish performance tests based on widely used benchmarks, and this time results released to coincide with each company's model announcement were compared.

The largest gap appeared in SWE-Bench Pro, a test that measures the ability to read large amounts of code for systems and fix bugs. In the test published by U.S. Scale AI, Mythos 5 reached an accuracy rate of 80.3%. By comparison, OpenAI's latest AI, GPT-5.5, scored 58.6%, while Google's Gemini 3.5 Flash scored 55.1%. In general, a gap of around 5 percentage points in accuracy is enough for performance differences to begin to show, and a gap of nearly 20 points is significant.

Mythos 5 also outperformed GPT-5.5 and Gemini in tests measuring the ability to operate browsers and apps, as well as the ability to build a system environment from a terminal. In a cyberattack-capability test conducted by the UK government's AI Security Institute (AISI), the previous-generation Mythos Preview only slightly outperformed GPT-5.5 as of mid-May, but this time the gap widened in its ability to handle complex tasks.

Mythos 5 stands out especially in long-duration work that would take humans 6 to 20 hours, or even several days. The ability to persist through such complex tasks could lead to greater precision in real-world cyberattacks, and it is considered an important indicator for assessing the danger of AI models.

Service suspension and the rise of domestic AI

Claude Fable 5, which added safety measures to Mythos and began general availability on the 9th, was also withdrawn on the 12th. By contrast, the availability of Mythos Preview, announced in April, has not been disclosed. However, Japan's three megabanks, Hitachi, and Trend Micro have obtained access rights, and some domestic companies have said, 'The preview is usable.'

The U.S. government's order to respond under the export control regime likely reflects the view that Mythos 5's cyber defense capabilities have reached a level far above OpenAI and Google. The government sees it as a potential security threat if misused by hostile countries.

Anthropic CEO Dario Amodei has said Chinese AI could catch up with Mythos in 6 to 12 months, underscoring the rapid pace of AI development. On the 22nd, OpenAI said it had updated GPT-5.5 Cyber, a model focused on cybersecurity capabilities, and that it slightly outperformed Mythos 5 in CyberGym, one of the tests used to measure cyberattack capability.

In Japan, Sakana Fugu, launched for general availability on the 22nd by Sakana AI in Minato, Tokyo, showed performance in programming that exceeded the latest AI from Google and others. Its feature is that multiple AI models can be switched and used, making continued use easier even if advanced AI from U.S. companies becomes subject to an import ban. As the suspension of U.S. AI services is prepared for, the importance of domestic AI is rising further.

U.S. government designates latest Anthropic AI for export controls, suspends access

Enjoyed this article? Share it with your network!