🇰🇷 Best AI for Korean
Use this one
- Best overall: Gemini 3.1 Pro (high) (84.1).
- Open-weight: none in this top 10.
- Caveat: Newest frontier models (e.g. GPT-6 Astra) have no Horangi 4 result yet. Same model can appear at different reasoning-effort settings.
- Translation: not separately ranked yet; see gaps.
| # | Model | Maker | Score ↑ |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (high) | 84.1 | |
| 2 | Claude Fable 5 (high) | Anthropic | 82.6 |
| 3 | GPT-5.5 (xhigh) | OpenAI | 81.8 |
| 4 | Gemini 3.5 Flash | 81.2 | |
| 5 | GPT-5.6 Sol (max) | OpenAI | 80.7 |
| 6 | Qwen3.7 Max | Alibaba | 80.2 |
| 7 | GPT-5.4 (xhigh) | OpenAI | 79.7 |
| 8 | Grok 4.5 | xAI | 78.7 |
| 9 | Claude Sonnet 5 (high) | Anthropic | 78.6 |
| 10 | GPT-5.6 Terra (max) | OpenAI | 78.3 |
Source: Horangi 4 (W&B Korea Korean LLM Leaderboard), via AIB (official report), data as of 2026-08-03, retrieved 2026-10-15. Metric: Final score %, mean of General Language Performance and Alignment over 20+ Korean tests, higher is better.
Also: KMMLU (Horangi 100-question subset): Gemini 3.1 Pro and GPT-5.6 Sol tie at 94.0%; GLM-5 is the top open-weight entry at 91.7%.
The smartest model isn't always deployable.
Get ahead of the constraints with The Signal.