South Korea Breaks Into US-China AI Top Tier

For the past two years, the absolute peak of the global artificial intelligence leaderboard has effectively been a duopoly. A Seoul-based startup now claims to have forced its way into that elite tier.

On May 30, Sung Kim, CEO of the South Korean firm Upstage, announced on Facebook that an unfinished version of his team's next-generation Solar model had crossed the 40-point threshold on the Artificial Analysis Intelligence Index, an independent benchmark that compresses a model's overall capabilities into a single score. According to Kim, only the United States and China had previously fielded models above that line. The claim, if verified, makes South Korea the third nation on a highly exclusive list—a feat achieved even before the model is fully completed, with the finished version slated for release at the end of June.

A Third Seat in the US-China AI Club

The rankings cited by Kim are compiled by Artificial Analysis, an independent firm that evaluates frontier models using a weighted average of approximately ten benchmarks spanning reasoning, mathematics, coding, and agentic capabilities. The benchmark is English- and text-based, and—crucially for such comparative claims—operates entirely independently of the AI laboratories it assesses.

The upper echelon of the index features familiar leaders: Anthropic's Claude Opus 4.8 at 61.4, OpenAI's GPT-5.5 at 60.2, and Google's Gemini 3.1 Pro Preview at 57.2, with Chinese models like Qwen and Kimi occupying the 50-point range. Upstage's model, listed as 'Solar-Preview,' has secured a spot in the 40s. The comparisons highlighted by Kim are particularly telling: the preview version outpaces Mistral Medium 3.5 at 39.2 points—Europe's strongest open model—and Canada-based Cohere's Command A+ at 37.2. In effect, the flagship models from Europe and Canada currently rank below this unfinished South Korean system.

The Pedigree of Upstage and the Solar LLM Line

For readers unfamiliar with the firm, Upstage is a six-year-old startup founded in 2020 by prominent researchers from Naver and Kakao, South Korea's leading internet conglomerates. The company initially built its reputation in document AI before expanding into its proprietary 'Solar' large language model (Solar LLM) family. Its breakthrough in 2024 centered on computing efficiency, debuting the Solar Pro model, which was optimized to run on a single GPU using its proprietary 'Depth-Up Scaling' methodology.

Since then, the product line has advanced steadily. Solar Pro 2 achieved near-frontier performance in 2025, while Solar Pro 3, launched this past March, approximately doubled the model's agentic capabilities while maintaining exceptionally high Korean-language output quality for its parameter size. The latest iteration reported to have cleared the 40-point threshold represents the company's next leap—a true foundation model developed entirely from the ground up rather than an incremental update.

Forged in the Sovereign-AI Survival Contest

To understand why Kim framed this achievement as a national milestone rather than a purely corporate one, it is necessary to examine the model's provenance. In August 2025, South Korea's Ministry of Science and ICT selected five consortia for its 'sovereign AI' initiative. Designed to build domestic frontier models and reduce reliance on U.S. and Chinese systems, the state-backed program provides designated developers with substantial GPU compute, specialized data, and top-tier talent. The initial cohort comprised Naver Cloud, LG AI Research, SK Telecom, NC AI, and Upstage.

Crucially, the initiative was structured as an elimination tournament, and the subsequent consolidation has been severe. By January 2026, two of the five participants—Naver Cloud and NC AI—were eliminated, leaving only SK Telecom, LG AI Research, and Upstage in the running. The Solar model that recently crossed the 40-point threshold was forged within this high-stakes competition, explaining why its performance is viewed in Seoul as a triumph for national technology strategy rather than a single enterprise.

Targeting 50 This Summer, 60 by Year-End

Kim does not view the 40-point milestone as a final destination. He disclosed that the team is targeting a score of 50-plus this summer, with the goal of reaching 60 by the end of the year—a threshold currently occupied exclusively by top-tier U.S. models. On the index chart he shared, the low 50s represents the territory of Meta-class systems, making the scale of Upstage's ambitions clear.

As is customary for founders in the midst of a competitive race, Kim concluded his announcement with a direct recruitment appeal, inviting engineers with LLM-building experience to join Upstage. For international observers, however, the immediate test is more straightforward: whether the finalized model slated for late June can sustain—or even exceed—the 40-point performance claim.