Alibaba has released API access to its flagship multimodal model Qwen4-Max, priced at $1.50 per million input tokens and $4.50 per million output tokens. The system is built on a 2.4-trillion parameter MoE architecture, supports a 1M token context window, and processes text, images, and video. According to Deep Learning, its combined reasoning and response limit reaches 262K tokens. According to Alibaba, the new model consumes half as many tokens as Qwen3.8-Max, reducing the cost per task execution by over 60%.

Compatibility and Integration

The programming interface supports OpenAI protocols—Chat Completions and Responses—as well as the Anthropic Messages specification. To switch agent wrappers to Alibaba's infrastructure, development teams only need to change the base URL, API key, and specify the qwen4-max model name. Compatibility with popular protocols lowers the migration barrier, sparing engineering teams from rewriting code for specific SDKs, though direct infrastructure expenses depend on actual token generation.

Performance and Hidden Costs

According to benchmarks by Artificial Analysis, the model scored 43 against a median of 41 in its price tier, with an estimated cost of $3.27 per task. However, during test runs, Qwen4-Max generated 127M tokens compared to a median of 78.6M. Because of this excessive generation volume, migrating enterprise agent systems to Qwen4-Max may prove more expensive than basic pricing calculations suggest: multiplying the elevated token consumption by rates of $4.50 per million output tokens means total expenses for large-scale autonomous tasks will noticeably exceed the nominal $3.27 per operation. Meanwhile, the flagship's weights remain closed, and the company has promised to release open smaller versions of the Qwen4 family later without specific deadlines.

The bottom line

Low API rates do not guarantee lower operational budgets if a model generates excessive token volumes during complex autonomous runs. Technical compatibility simplifies migration, but architectural appetite determines actual unit economics.

What came before

Why would engineering teams rewire production pipelines for a nominally cheap API? While in late April the company launched a compact model with just 27 billion parameters to question bloated systems, Alibaba now pivots straight back to a massive 2.4-trillion-parameter architecture to win the price sheet.

Our stories: Qwen3.6-27B vs MoE: How Efficiency Beats Model Size in AI Coding (April 26).

Large Language ModelsEnterprise AI APIsModel InferenceCost Reduction