Qwen3.8 Max matches Claude Opus 4.8 but trails Kimi K3 at higher cost

Alibaba released Qwen3.8 Max on the Artificial Analysis Intelligence Index with a score of 56, marking a significant 10-point improvement over Qwen3.7 Max (46). According to Artificial Analysis, this puts Qwen3.8 Max on par with Claude Opus 4.8 and ahead of GLM-5.2 (51), but it falls short of Kimi K3 (57)—the higher-scoring alternative that also runs 25 percent cheaper.

On GDPval-AA, a benchmark designed to measure work-related task performance, Qwen3.8 Max jumped 468 Elo points to 1,739, surpassing Kimi K3 (1,685) but remaining below Claude Opus 5 (1,852). However, the method behind this improvement reveals a costly trade-off. The model achieves higher scores by employing 64 steps per task instead of 14, and it resends the full conversation history to the model at each step, causing input tokens to grow 15×.

This architectural choice has direct economic consequences. Although Alibaba reduced token prices—input dropped from $2.50 to $2.00 per million tokens, output from $7.50 to $6.00 per million tokens, and cache hits from $0.50 to $0.25—the volume increase means a single task on the Intelligence Index now costs $1.14, more than double Qwen3.7 Max ($0.53). By contrast, Kimi K3 scores one point higher at just $0.86 per task, and GLM-5.2 comes in at $0.57, making both substantially more cost-effective.

Qwen3.8 Max also exhibits concerning regressions. AA-LCR, which tests whether a model can correctly extract information from very long texts, dropped 2 points. AA-Omniscience, measuring accurate knowledge responses or honest admission of uncertainty, fell 10 points. The accuracy rate holds around 31 percent, but the hallucination rate jumped from 23 to 40 percent—the model guesses far more often instead of admitting it does not know, a behavioral shift that may undermine reliability despite surface-level performance gains.