Qwen4-Max Cuts Average Task Cost 62% to $1.67
The November 11 preview was Token Plan only. The pay-as-you-go API is now live, and the list price sits 25% below Qwen3.8-Max.
Published 5 min read

Alibaba's Qwen team opened Qwen4-Max to general access on November 16, 2026, at 05:01 Turkey time (10:01 China time; 18:01 PT on November 15). The model is live on the QwenCloud API, on the Token Plan, and in Qwen Studio, QwenWork, and Qoder. The earlier qwen4-max-preview, which opened on November 11, was Token Plan only and did not support pay-as-you-go.
Qwen announced the launch on X:
Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
November 16, 2026
In its launch blog, Qwen says the model is built on the new Qwen4 architecture. It has a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, and an additional 328 billion N-gram Embedding parameters. It has native vision and a 1,000,000-token context window.
Why does the launch matter now?
The news is not only a new model name. The API is now pay-as-you-go, and the list price is lower. On the Alibaba Cloud Model Studio pricing page, qwen4-max is $1.50 per million input tokens and $4.50 per million output tokens on the Singapore international list. Qwen3.8-Max is $2 and $6 in the same table. That is a 25% lower list price on both input and output.
Implicit cache reads cost $0.075 per million tokens. Qwen says that line is 70% lower than on Qwen3.8-Max. Explicit cache writes are $1.875 for the 5-minute option and $3 for the 1-hour option. Reads are also $0.075.
In Qwen's own Work360 evaluation, the average cost per task falls from $4.43 to $1.67, a 62% drop. The X post summarized the same internal test as "up to ~50% fewer tokens" and "average cost per task drops over 60%." The blog gives the more precise figures.
What does the new architecture change?
Qwen4-Max is a mixture-of-experts model. The blog says the team updated attention, residual connections, and embeddings. The three named pieces are GDN + Qwen Sparse Attention (QSA), Gated Residual (GR), and N-gram Embedding.
QwenCloud documentation lists a maximum output of 131,072 tokens in non-thinking mode. With thinking on, thinking plus the final answer can reach 262,144 tokens. The default thinking budget is 131,072 tokens. Thinking is on by default, and the Chat Completions docs set the default effort to high.
The model accepts text, images, and video. It supports function calling, structured output, web search, PDFs, a code interpreter, and MCP (Model Context Protocol). Fine-tuning and model deployment are not supported. Batch inference is also not supported.
What do Qwen's tests show?
Work360 is Qwen's framework for complex work scenarios. It spans 26 Tier-1 domains, 182 Tier-2 subdomains, and 721 Tier-3 specialties. In that internal test, Qwen4-Max scored 52.69, and Qwen3.8-Max scored 47.74.
Token use also falls, again on Qwen's measurements. On Work360, usage drops by about 24% at Xhigh effort and about 35% at High effort. On Terminal-Bench 4.0, the reductions are about 48% and 49%. These are the company's own figures.
The announcement graphic compares Qwen4-Max with Qwen3.8-Max-0902, Opus 5.5, GPT-6.1-Sol, GPT-6 Astra, and Kimi K3.1. The new model is not first on every test. On Qwen's slide, it scores 75.7 on DeepSWE v1.1, up from 69.3, 60.0 on HealthBench Professional, up from 39.6, and 21.3 on HLAB-Vals-1.1. On Terminal-Bench 4.0 it scores 48.5, while Kimi K3.1 is at 67.2 and Opus 5.5 at 59.6.
The same day, Qwen posted follow-ups from its own account. It said Qwen4-Max reached sixth place in Code Arena WebDev, 74 points above Qwen3.8-Max. It also said the model posted a 7.5% net gain over the previous Max on Agent Arena, with Alibaba fifth among labs overall. In Image-to-WebDev Arena it was sixth with 1,674 points, 35 points above the previous Max at 1,639. On DesignArena it was fourth in Python-PPTX and fifth in Spreadsheet. The model is also live on OpenRouter. Those notes come from Qwen's own posts.
Where can developers use it?
The QwenCloud region list does not include Turkey or Istanbul. Frankfurt is the European option. Other listed regions are Singapore, Hong Kong, Tokyo, Beijing, and US Virginia. The docs say Beijing and Global share prices, while Singapore has its own.
The API is compatible with OpenAI Chat Completions, Responses, Anthropic Messages, and DashScope, so many existing developer tools can point at it. Pay-as-you-go calls in Singapore include 1 million free tokens, separate from Token Plan credits.
Qwen4-Max itself is not open weights. There is no Hugging Face or GitHub weight release for this model. The blog and the X post say smaller Qwen4 models will be released as open weights. They do not give sizes or a date. Fine-tuning and deployment are not supported, so this flagship is not a self-hosted model.
The blog also includes manufacturing, legal, finance, and lab showcases. In the SANF and Pacific Union examples, Qwen says programming time for a single workpiece fell from about 475 minutes to 24 minutes, a drop of about 95%. Those are company demonstrations, not independent measurements.
Frequently asked questions
Is Qwen4-Max free?
No. It is a pay-as-you-go API. Singapore offers a 1 million-token free quota. The Token Plan is a separate subscription.
How is this different from the November 11 preview?
qwen4-max-preview was Token Plan only. That preview runs from 10:00 on November 11 to 24:00 on November 25 (UTC+8). The qwen4-max model that opened on November 16 is a separate model ID with pay-as-you-go access.
Can developers in Turkey use it?
There is no Turkey region on the list. Frankfurt and Singapore are listed options. The API is compatible with OpenAI and Anthropic formats.
Is Qwen4-Max open weights?
No. Qwen says smaller Qwen4 models will be released as open weights later. It has not given a date.