What is the best LLM
across languages?
Ranked from two independent public multilingual suites: Artificial Analysis Global-MMLU-Lite and the AI Language Proficiency Monitor. The general ranking uses one aggregate row per source. It does not let ten language columns count as ten separate votes. A named winner also requires 20 callable models with current two-source coverage.
No current winner is published: qualifying independent evidence is older than 30 days. The table below shows available corroborated evidence, not a publishable current winner.
| # | Model | Multilingual | Reasoning | Price/M |
|---|---|---|---|---|
| 1 | Anthropic: Claude Opus 4.6 | 95.0 | - | $5.00/M |
| 2 | Anthropic: Claude Sonnet 4.6 | 93.7 | 67.9 | $3.00/M |
| 3 | Anthropic: Claude Opus 4.5 | 93.4 | 71.4 | $5.00/M |
| 4 | OpenAI: GPT-5.2 | 92.2 | 67.6 | $1.75/M |
| 5 | Anthropic: Claude Sonnet 4.5 | 92.1 | 43.1 | $3.00/M |
| 6 | Google: Gemini 2.5 Pro | 92.0 | 53.2 | $1.25/M |
| 7 | Anthropic: Claude Haiku 4.5 | 88.7 | - | $1.00/M |
| 8 | Meta: Llama 4 Maverick | 86.4 | 52.2 | $0.15/M |
| 9 | Meta: Llama 3.3 70B Instruct | 81.0 | 60.4 | $0.10/M |
| 10 | Meta: Llama 4 Scout | 77.1 | - | $0.10/M |
| 11 | Google: Gemini 2.5 Flash | 58.6 | 29.2 | $0.30/M |
The Language Proficiency Monitor row averages Arabic, Bengali, German, Spanish, French, Hindi, Japanese, Portuguese, Russian, and Mandarin Chinese.
Each included model must have the full ten-language, three-task panel. Missing tests do not quietly produce an inflated average.
Per-language scores are stored for the next niche pages. They are not duplicate weight in this broad ranking.
Each page computes its own two-source composite. Russian is withheld until a second independent per-language source is available.
What is the best multilingual LLM?
No current winner is published: qualifying independent evidence is older than 30 days.
How is the multilingual ranking calculated?
The composite uses one aggregate score from Artificial Analysis Global-MMLU-Lite and one from the independently published AI Language Proficiency Monitor. Per-language rows are kept for language-specific rankings, not multiplied into this broad score.