Qwen4-Max lowers its prices and targets long-duration agents
Alibaba Cloud makes Qwen4-Max generally available with a context of one million tokens and reduced listed prices. The model mainly targets agentic tasks, but the bill still depends on reasoning, outputs, and caching.
Alibaba Cloud made Qwen4-Max generally available on November 16, 2026, via QwenCloud, following an open preview phase starting November 11. This multimodal model, featuring a context window of up to one million tokens, arrives with listed prices lower than those of Qwen3.8-Max and a clear focus on agent, coding, and productivity use cases. On paper, the price reduction is real; in practice, the final cost still heavily depends on response length and reuse of inputs.
A flagship model betting on long context
Qwen4-Max is based on a mixture-of-experts architecture and features 2.4 trillion parameters, according to Alibaba. The model directly accepts images, supports text output, and can handle more complex workflows thanks to streaming, web search, parallel tool calls, functions, prefix pre-filling, and structured JSON output. In other words, it not only responds but also chains multiple actions without losing track too quickly.
Alibaba presents it for coding, office productivity, front-end design, and complex agentic tasks. The reasoning mode is enabled by default at a high level, with low, high, and xhigh settings. A useful fact to keep in mind: the thinking budget and the final response share the same ceiling of 262,144 tokens, meaning that deliberation eats into output capacity.
Prices lowered, but total cost not
In Singapore, Alibaba Cloud lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens, compared to $2 and $6 for Qwen3.8-Max. This is a 25% reduction on both pricing lines. A free allocation of one million tokens is also offered for 90 days from the latest of the following three dates: activation of Model Studio, model publication, or request approval.
However, the economic calculation does not stop there. Alibaba states that the model consumes fewer tokens than Qwen3.8-Max on comparable tasks and claims to have reduced the average task cost on its internal Work360 benchmark from 26.58 yuan to 11.16 yuan, a decrease of about 58%. This difference is due both to the face price and the model's efficiency. For teams working on long agents, the bill depends on three variables: input volume, output length, and reasoning usage.
Cache becomes a real lever for repetitive workflows
Alibaba Cloud also details an explicit cache system that can reduce the cost of identical prefixes, such as system prompts, tool descriptions, or reference documents. The explicit cache requires at least 1,024 tokens and can last five minutes or one hour. Cache reads cost $0.075 per million tokens in Singapore, while creating an explicit cache costs $1.875 for five minutes and $3 for one hour.
Implicit caches automatically detect common prefixes without guaranteeing results. This mechanism mainly applies to agents that reuse long instructions from one call to another. In this context, prefix stability becomes almost a high-level sport.
Access, restrictions, and platform strategy
Qwen4-Max is listed in Beijing, Singapore, Frankfurt, Virginia, Tokyo, and Hong Kong. Among these regions, batch inference is only available in Beijing, while fine-tuning and custom model deployment remain unavailable everywhere in the provided documentation. The model is also integrated into Singapore's Personal Edition Token Plan, which operates via credits and supports Claude Code, Cursor, Qwen Code, Qoder, and OpenClaw.
This plan is not universal access: it prohibits automation scripts, custom application backends, and non-interactive batch calls. For these uses, the pay-as-you-go API must be used. The Qwen4-Max preview version, launched on November 11, 2026, at 10:00 UTC+8, was scheduled to run until November 25, 2026, before being withdrawn.
By the numbers
- 2.4 trillion parameters — model architecture, according to Alibaba.
- 1 million tokens — announced maximum context window.
- 25% — reduction in listed prices in Singapore compared to Qwen3.8-Max.
- 58% — decrease in average cost on Alibaba's Work360 benchmark.
- 90 days — duration of the free allocation announced in Singapore.