Alibaba's Qwen4-Max is available through its cloud service at $1.50 per million input tokens and $4.50 per million output tokens, which is 25% below the model it replaces. Running it through a standard set of tasks, though, costs $3.27 a task instead of $10.02, a fall of 67%. A cut of a quarter in the price cannot explain a cut of two thirds in the bill, and the gap between them is the useful part.
The bill has three parts, and the price list shows one
The cost of an AI job is the price of a token, times the number of tokens the model writes while it thinks and works, plus the cost of the text it reads. An agent sends the whole conversation so far back to the model at every step, so repeated text is billed at a discount. Qwen4-Max improved all three at once: tokens cost 25% less, repeated text costs 70% less, and the model needs about 105,000 output tokens per task where its predecessor needed 171,000.
Take output alone. At the old price of $6 per million tokens, 171,000 tokens cost about $1.03. At $4.50, 105,000 tokens cost about $0.47. A 25% price cut and 39% fewer tokens together give a bill a little under half as large, and the cheaper input and repeated text account for the rest of the fall.
The model also reads long documents with less effort. Three of its layers squeeze the earlier text into a short summary and a fourth picks out the parts that matter, so it no longer gives every word the same attention. For an agent that rereads a growing pile of text at each step, that is where the cost builds up.
The same gap shows up against rivals
Moonshot's Kimi K3.1 Pro charges $2 and $10 per million tokens, more than twice Qwen4-Max's output price. On the same set of tasks it costs $3.68 against Qwen's $3.27, so Qwen is only 11% cheaper, and Kimi scores three points higher on the index. Compared by price list, Qwen4-Max looks like a bargain; compared by finished work, it is a close call.
Cheapest per task is not the answer either. Tencent's Hy4 costs $1.03 a task but scores eight points lower, and a task done badly has to be done again. What a business wants is the cost of a task done right, which no pricing page prints.
Mistakes are the fourth cost
A run that misreads the brief costs as much as one that gets it right, and with long agent runs the error is often found hours later. Qwen4-Max comes with a feature called Live Delivery: the agent hands over a usable first version early and improves it in later passes, and the person can steer it on the way. Alibaba has released the plugin as open source.
For example, a finance team asking an agent for a month-end variance report could see in the first draft that a cost centre was grouped wrongly, and fix it then. That saves the whole run, which is worth more than a few percent off the token price.
Qwen4-Max is a closed model that you reach through Alibaba's API; only smaller Qwen4 models are promised as open weights later. Whichever model you look at, the habit that holds up when prices change is the same. Pick three jobs your team repeats, run each one through two or three models, and write down what each run cost and whether the result was usable.