More AI service companies are building in-house models to cope with the cost of using advanced AI models and to respond to competition with large model developers.

Legal AI startup Harvey has recently joined the push to build its own model. Harvey developed the Tenet model based on Chinese startup Moonshot's low-cost open-source model, Kimi K3. Harvey is expected to use Tenet for tasks it can cover, and use frontier models such as OpenAI or Anthropic when needed.

Coding AI Cursor is also accelerating its own model development. It launched Composer 2, an AI model with high cost efficiency, in March. After its acquisition by SpaceX was recently completed, there has also been talk it will soon unveil a new model developed in cooperation with SpaceX. Cursor is also expanding its AI development environment by launching its own source code hosting platform, Origin.

• Legal AI startup Harvey unveils its in-house legal model Tenet, foreshadowing upheaval in the legal market • Cursor launches AI coding model Composer 2, optimized for programming • SpaceX AI and Cursor said to be close to releasing a jointly developed AI model • Cursor unveils its own code hosting platform Origin, challenging GitHub

An analysis says these moves are not unrelated to the faster push by OpenAI and Anthropic to develop AI services based on their own AI models.

• [Tech Inside] "OpenAI and Anthropic will increasingly restrict access to advanced model APIs"... why?

In the second evaluation results for the Ministry of Science and ICT's Dokpamo project, Motif Technologies was eliminated while Upstage, SK Telecom and LG AI Research advanced to the third evaluation round. The government is pushing to revamp the way it supports selected companies from the third evaluation.

• In Dokpamo's second round, usability determined who advanced; it will be reshaped into 'frontier AI' after the final 2 teams are selected • Government pushes to revamp the Dokpamo project format • Up to multimodal and agents... SKT moves to scale up its proprietary AI • LG AI Research: "We passed Dokpamo's second round and secured a basis to compete with global frontier models"

Interest in physical AI appears to be rising again, symbolised by China's humanoid robot company Unitree. Wang Xingxing (왕싱싱), Unitree's chief executive officer, said at the Beijing World Robot Conference that under ideal conditions such a turning point could come in 2 to 3 years, and in 5 to 10 years if progress is slow. Amazon Web Services will introduce in South Korea a physical AI development program called the AWS Korea Physical AI Frontier Program.

• Unitree surges 629 percent on the first day of its Shanghai STAR listing • China's Unitree unveils new humanoid Superman, jumping 2 meters in place and running at 45.6 kph • Unitree CEO: "A turning point for humanoid robots is not far off... at the latest within 10 years" • AWS to launch its Physical AI Frontier Program in South Korea

It also summarised moves by domestic and overseas companies around AI.

OpenAI recently held a "ChatGPT Work media demo" and introduced key features and real-world use cases of ChatGPT Work. At the event, details were shared on how reporters and researchers are attaching Codex to their existing work. OpenAI also launched ChatGPT for Teens, which applies age-appropriate learning and safety features by default for teenage users.

• ChatGPT Work and Codex: how have they changed white-collar work? • OpenAI launches ChatGPT for teens, strengthening age-tailored guardrails

Adobe officially launched audio features for its all-in-one creative AI studio, Adobe Firefly. Salesforce released Slack Code, which lets teams call up coding agents in Slack to work together and share results. Slack Code supports interaction with coding agents within chat, and lets the entire team review generated code and change histories in the same context. AI-native creator marketing automation startup Storika introduced an open beta of an AI agentic platform that automates the entire process of brand creator marketing.

• Adobe Firefly officially launches music, voice and sound effect generation features • Storika launches creator marketing automation platform based on AI agents

Salesforce introduced Headless Data 360 for MCP so AI agents can use customer data directly. Matthew Miller (매튜 밀러), vice president for global Tableau product management at Salesforce, attended Salesforce Korea's annual AI and data conference, Datafam Seoul 2026, and stressed that it is now the era of "agentic analytics," in which insights find users first rather than users searching for insights. Miller said analysis in the past worked by people directly using tools such as Tableau to understand data. AI agents, by contrast, identify what users need and remember how they work. AI agents can also adapt themselves to the different terminology and ways of thinking of each organisation.

• Salesforce unveils Slack Code, a team collaboration coding agent feature in Slack • Salesforce unveils Headless Data 360, enabling AI agents to use customer data directly • "Analytics is also entering the agentic era... AI tells you insights first"

Snowflake added dynamic model routing to its Cortex AI Gateway.

• Snowflake adds dynamic model routing to Cortex AI Gateway, cutting costs by up to 3 times

MegazoneSoft completed Nongshim Group's project to build an AX, or AI transformation, system based on Google Workspace. The company said the project went beyond a simple system rollout. It included migrating existing data to Google Drive, linking legacy systems and third-party solutions with Google Workspace, and change management to reorganise group-wide work practices around AI. MegazoneCloud joined the Partner Agent Factory, an AWS program for joint development of AI agent solutions, and jointly developed 3 enterprise AI agent solutions with AWS.

• MegazoneSoft builds Nongshim Group AX system based on Google Workspace • MegazoneCloud joins AWS Partner Agent Factory and develops 3 enterprise AI agents

As companies across a range of fields move into AX support businesses, adult education content company Day1Company is also stepping up its AX business that supports corporate AI build and deployment processes, called the DAY1 AI Deployment Company. NCOA signed a strategic business agreement with Hanwha Systems to expand AX projects in the manufacturing sector. The companies plan to cooperate on designing and building an ontology-based knowledge system for manufacturing AX projects, implementing services based on knowledge graphs, establishing a data governance system for AI use, building a platform using NCOA's Ready Data solution, and jointly identifying and carrying out AX projects. AI services and solutions company BespinGlobal was selected as an OpenAI Select Partner in the OpenAI Partner Network.

• A surge of companies jumping into AX business; education firm Day1Company also enters • NCOA cooperates with Hanwha Systems on manufacturing AX, jointly developing industry-specific AI services • BespinGlobal selected as an OpenAI Select Partner, strengthening its corporate AI push

Google Cloud said Shinhan Securities worked with Google Cloud to build an AI agent platform linked to key internal business data and systems while complying with financial security regulations. Alibaba Cloud officially opened its third data center in South Korea, targeting the AI market.

• Shinhan Securities works with Google Cloud to build its own AI agent platform • Alibaba Cloud opens its third data center in South Korea, targeting the enterprise agentic AI market

Major companies' strategies are taking shape as they join the government's "AI for Everyone" project. They share a focus on general-purpose chatbots, but their strategies diverge on which models to combine and how to connect them with existing services.

• "AI for Everyone" bids: SKT full stack, Kakao for everyday touchpoints, KT multi-LLM

A revised Personal Information Protection Act containing special provisions for the use of personal information to develop AI technology passed the National Assembly plenary session on the 20th. Under the change, a special provision was 마련됐다 that allows legally collected personal information to be used for AI technology development, on the condition that AI development is difficult with only pseudonymised or anonymised information, that public interest and social necessity are recognised, and that strengthened safety measures and deliberation and a resolution by the Personal Information Protection Commission are in place.

• Amendment to the Personal Information Protection Act passes, introducing special provisions for personal information use in AI

With concerns that AI-generated content is flooding the internet and polluting the information ecosystem, tech platform companies are moving to carry out a large-scale cleanup.

• War on AI spam 'slop'... tech platforms launch a major cleanup