Atlas

Benchmarks

← All benchmarks

Arena-Hard-Auto v2.0 - Creative Writing

Chat & Writing · 2025-04-23

The 250 creative-writing prompts of Arena-Hard-Auto v2.0, scored as the estimated win rate in percent against the v2.0 creative-writing baseline under an LLM judge. A component of the Arena-Hard v2.0 family (arena-hard-v2), measured on the same runs as the hard-prompts subset. Higher is better.

Top models (higher is better)

ModelScore
Kimi K2 Instruct85.0
Nemotron Cascade 2 30B A3B78.7
Opus 468.5
Qwen3.5 35B A3B66.3
Qwen3 30B A3B Thinking 250766.0
Nemotron 3 Nano 30B A3B63.2
GPT-4.161.5
Qwen3-235B-A22B59.8
DeepSeek-V3-032459.3
Sonnet 454.6
gpt-oss-20b25.9
Loading Atlas data…