Anthropic opened a Seoul office and announced a series of partnerships across South Korea’s artificial intelligence ecosystem as the Claude developer expands its presence in Asia. As the company prepares for a public listing, Anthropic identified South Korea as one of Claude’s most active markets globally, with usage concentrated in technical and creative work. The international expansion comes as Anthropic seeks to restore access to some of its most advanced AI models following a new export-control order.

Public and Private Sector Partnerships

Anthropic signed a memorandum of understanding with South Korea’s Ministry of Science and ICT to collaborate on public-sector AI adoption, model safety testing, and efforts to address AI-related cybersecurity threats. Major South Korean business groups are deploying Claude through affiliates including LG CNS, Hanwha Solutions, and Samsung SDS. Employees at these firms are using the AI assistant for software development, knowledge work, and enterprise operations.

Internet giant NAVER is using Claude Code to accelerate software development, while Nexon has deployed Claude across game development workflows.

Academic and Startup Initiatives

Anthropic will work with South Korea’s National AI Research Lab to provide Claude access to about 60 researchers working on AI safety, model evaluation, alignment, and frontier AI research. As part of its expansion, the company is holding its first Seoul Builder Summit with AI startup Coxwave, aimed at developers, startups, and businesses using Claude. The ‘Claude for Startups’ program is currently live in the country, providing early-stage companies with model credits, technical support, and access to Anthropic’s AI tools.

The Seoul office will be headed by Kiyoung Choi, who joined Anthropic in May as Representative Director of Korea from Snowflake Inc.

Anthropic has released findings from its "Project Fetch" Phase Two, indicating that its Claude Opus 4.7 model can outperform human teams in robotics planning tasks. The company reported that the model completed tested objectives at speeds roughly 20 times faster than the quickest human teams from the previous year. While the model demonstrated significant speed advantages, it faced limitations in precision control, highlighting a gap between planning and physical execution.

Performance Comparison

The trials involved three tests using Claude Opus 4.7 within Claude Code. A researcher connected a laptop to the robot, entered the initial prompt, and approved commands and task transitions. The results were compared against an original experiment from August 2025, which pitted Anthropic employees using Claude support against a group limited to web research and their own problem-solving.

Capabilities and Limitations

Anthropic noted that Opus 4.7 moved quickly through choices that typically slow humans down, such as interfacing with the robot's sensors. The model produced less code than the Claude-assisted human team while matching or exceeding outcomes on the tested tasks. However, the company emphasized that these results do not constitute a robotics breakthrough. The model struggled with the "fetching" portion of the test, which required precisely guiding a beach ball back to a start area using environmental feedback. While Opus 4.7 could position the robot to attempt the push, the motion control was not accurate enough to complete the task.

Future Research

The improvements were driven by broader model scaling rather than targeted robotics research. Anthropic stated that more research is needed to understand the models' ability to make physical tools more bespoke, whether by writing control policies or designing robotic systems. The company suggested that while barriers exist to generalized physically capable language models, rapid advancements in software tools suggest a similar trajectory in hardware is possible.