Qwen3.8 Max catches Claude Opus 4.8 but Kimi K3 still scores higher for 25 percent less

Why it matters

Alibaba's Qwen3.8 Max posts a strong gain on the Artificial Analysis Intelligence Index, drawing even with Claude Opus 4.8 and overtaking GLM-5.2, yet it still trails Kimi K3, which is both cheaper per task and slightly higher scoring. On the work-task benchmark GDPval-AA, the new model passes Kimi K3 and sits behind only Claude Opus 5. The improvement comes at a price: the model takes far more steps per task and processes much larger inputs, more than doubling per-task cost compared with Qwen3.7 Max despite lower token pricing. Beyond cost, there are quality regressions. Performance on long-context reasoning slipped slightly, and AA-Omniscience dropped, with the hallucination rate climbing sharply while accuracy held steady, meaning the model guesses rather than admits uncertainty. For readers, the story illustrates that headline benchmark gains can mask tradeoffs in efficiency, cost, and reliability, and that rivals can deliver comparable or better results at lower expense.