Alibaba removed Qwen4-MAX from the preview: the model appeared in the API with a payment for tokens, and #036;1.50 for 1 million input tokens and #036; 4.5 for the weekend, on OpenRouter the price is the same. In the preview I wrote about it, it was only granted by subscribing to the ken program. Artificial analysis has already added the model to its index: 43 points. That's above DeepSeek V4.1 Flash33 and GLM-5.3, below Kimi K3.1 Pro and far behind GPT-6.1 Sol (55), Claude Opus 5.5 (59) and Claude Fable 5.2. It is not cheap at the same time: Qwen4-MAX is the operator, and the average task index is & #036;3.27. Gemini 4 Hydrogen at x High Score 54 points for $3.16. Alibaba himself discovered the size: MoE on 2.4 trillion parameters, 71 billion active to mark, the context 1 million. Qwen4-MAX is the first in the latest test of the agent and HLAB-Vals, while Terminal-Bench 4.0 is 48.5 compared to 59.6 for Opus 5.5. No scales, they promise to unlock the young Qwen4 models. Colleagues discussed more on the @neuro channel.