Qwen3.8 Max, Claude Opus와 대등 — 하지만 할루시네이션 74% 급증
Alibaba의 Qwen3.8 Max가 벤치마크에서 Claude Opus 4.8과 대등한 성능을 보이지만, 할루시네이션 비율이 23%에서 40%로 급증하며 성능과 신뢰성 사이의 중대한 트레이드오프를 시사하고 있다.
alibaba released qwen3.8 max, achieving a score of 56 on the artificial analysis index to match claude opus 4.8, while also jumping 468 elo points on the gdpval-aa benchmark to 1,739. despite these gains, the model requires 64 steps per task and costs $1.14 per task—more than double its predecessor—while competitor kimi k3 scores higher for less money. furthermore, benchmarks show regressions: aa-omniscience fell 10 points, and the hallucination rate surged from 23% to 40%, meaning the model guesses more often rather than admitting ignorance. this indicates that while raw capability increased, the reliability necessary for enterprise deployment may have compromised.