Alibaba released Qwen4-Max on November 16, opening paid API access to a flagship model with native visual understanding and a one-million-token context window. The launch follows a separate Token Plan preview introduced on November 11. QwenCloud’s dated release notes list the production model as qwen4-max, alongside support for tool calls, structured output and caching.
API prices depend on the deployment region
QwenCloud lists Singapore pricing at $1.50 per million input tokens and $4.50 per million output tokens. Beijing and Global deployments cost $1.413 and $4.24 respectively. Global endpoints include Frankfurt, Hong Kong, Tokyo and Virginia.
Thinking and non-thinking modes use the same rates, but reasoning tokens count toward the output bill. Qwen4-Max has no input-length price tiers. Cached reads cost $0.075 per million tokens in Singapore, with separate charges for writing a cache. A low cached-input rate therefore describes one part of the bill, rather than the cost of an entire agent run.
Tools and reusable context support longer tasks
The developer guide supports image and video understanding, parallel tool calls and MCP, alongside OpenAI-compatible and Anthropic-compatible interfaces. That gives builders of coding agents another hosted model to evaluate within existing tool workflows.
Explicit caching can reuse shared prompts, tool definitions and reference documents for five minutes or one hour. Support depends on the API: the guide excludes the DashScope interface from the new explicit-cache feature. Batch inference, fine-tuning and model deployment are also listed as unsupported for Qwen4-Max.
Alibaba’s launch says smaller open-weight Qwen4 models will follow. That is a future release plan; the flagship’s API launch does not make its weights available.
Efficiency figures come from Alibaba’s own evaluation
Alibaba says Qwen4-Max has 2.4 trillion total parameters, with 71 billion activated per token. On its proprietary Work360 benchmark, it reports roughly 60% lower average task cost and 35% fewer tokens than Qwen3.8-Max at High reasoning effort.
Those are vendor-reported results on a particular workload, not a guaranteed reduction in a customer’s bill. The release also describes extended scientific workflows, including a 48-hour agent run. For teams evaluating agent infrastructure, the useful comparison is whether the model completes their actual tasks reliably, with the required tools and regional endpoints, at an acceptable total cost.