A Flagship That Costs Less Than Its Predecessor
Alibaba's Qwen team introduced Qwen4-Max on 16 November 2026: 2.4 trillion parameters in total, 71 billion active per token, native vision and a one-million-token context window. Alibaba Cloud's Model Studio lists it at US$1.50 per million input tokens and US$4.50 per million output tokens in Singapore, and slightly less in its other regions. Artificial Analysis lists Qwen3.8-Max at US$2 and US$6, so the new model is 25 percent cheaper on list price.
| Model and region | Input per million tokens | Output per million tokens |
|---|---|---|
| Qwen4-Max, Singapore | US$1.50 | US$4.50 |
| Qwen4-Max, Frankfurt and other regions | US$1.413 | US$4.24 |
| Qwen3.8-Max (Artificial Analysis listing) | US$2.00 | US$6.00 |
What Alibaba Claims and What It Has Not Shown
Alibaba says the average cost per task on its own Work360 benchmark fell from US$4.43 to US$1.67, a 62 percent cut, and that the Work360 score rose from 47.74 to 52.69. Both figures come from a benchmark Alibaba designed, so treat them as a hypothesis. The model is also not open-weight: Alibaba says only smaller Qwen4 models will be open-sourced later, which means you rent this one and cannot host it yourself.
What to Do Before You Move Traffic
Take 20 real tasks from your production traffic, run them on your current model and on Qwen4-Max, and compare the cost per finished task, not the price per token. The API accepts OpenAI and Anthropic message formats, so the test needs little code. The documentation lists Frankfurt among its global regions, so ask Alibaba in writing where prompts and logs are processed and stored before you send any personal data.
Read next: The Qwen That Needed 2.4 Terabytes Just Shipped One That Fits on a GPU | Qwen's 3 Billion Downloads Hide a Compliance Gap



