Benchmark

Arena-Hard Benchmark

What is Arena-Hard?

LLM-judge dependent, with known style and length bias. Benchmark Radar tracks 3 reported scores for Arena-Hard, cited by 4 source documents, with links to its…

Released: 2024-04-19
Evaluates: human preference
Openness: unknown
Importer: Model reports
Reported scores: 3

Source provenance

Original evidence https://github.com/lmarena/arena-hard-auto

Arena-Hard code

Code https://github.com/lmarena/arena-hard-auto

Arena-Hard results and reported scores

Model reports

Model	Organization	Reported value	Reported	Evidence
DeepSeek-R1	DeepSeek	92.3	2025-01-22	evidence
DeepSeek-V3	DeepSeek	85.5	2024-12-27	evidence
Qwen3-235B-A22B (Thinking)	Qwen	95.6	2025-05-14	evidence

Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.

Which sources cite Arena-Hard?

Qwen3 Qwen, 2025-05-14
DeepSeek-R1 DeepSeek, 2025-01-22
DeepSeek-V3 DeepSeek, 2024-12-27
Llama 3.1 405B Meta, 2024-07-23