Source: the-decoder.com/alibabas-qwen4-max-takes-on-opus-5-5-gpt-6-and-kimi-k3-1-pro-at-much-lower-prices
Alibaba has released Qwen4-Max, a mixture-of-experts model with 2.4 trillion parameters, of which 71 billion are active per token, according to the Qwen team. It processes images natively and offers a one-million-token context window. Alibaba says a hybrid attention design compresses earlier context and then picks out the sections that matter, which reduces compute on long inputs.
On the Artificial Analysis Intelligence Index, Qwen4-Max scores 43, nine points above its predecessor Qwen3.8-Max and three behind Kimi K3.1 Pro at 46. Artificial Analysis puts its cost at $3.27 per task, against $10.02 for Qwen3.8-Max. Through Alibaba's own API, input tokens cost $1.50 per million and output tokens $4.50 per million, with cache hits at $0.075. On Terminal-Bench 4.0 it scores 48 percent, behind Kimi K3.1 Pro at 67 percent.
Alibaba's Live Delivery feature provides a usable first version early and lets users inspect intermediate results and change course while long-running agents work. The company has open-sourced the accompanying Qwen-Live-Delivery-Plugin. Qwen4-Max is available through QwenCloud, QwenWork, Qoder and QwenStudio, and Alibaba plans to release smaller Qwen4 models as open source later.
Why it matters if you build agents
Developers running long agent tasks get a lower-priced option with a large context window and a workflow where they can step in mid-run. The benchmark results are mixed and largely come from Alibaba and Artificial Analysis, and Kimi K3.1 Pro leads on several agent tests. Only the smaller Qwen4 models will get open weights for now, so the flagship has to be used through Alibaba's hosted services.
Get stories like this every weekday
One short email with the day's agent tooling news.