Alibaba launches Qwen4-Max: a 1M-token model for coding and agents, priced from $1.41
Alibaba's Qwen4-Max is live in Model Studio at about $1.41–$1.50 per million input tokens, with a 1M-token context and OpenAI- and Anthropic-compatible APIs. What it does and what to check.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 4 understood

Alibaba announced Qwen4-Max on 16 November 2026, calling it the most capable model in its Qwen family so far. It is a large aimed at coding, agent tasks and office work, and it is already listed with prices in Alibaba Cloud Model Studio, including regions in Frankfurt and Virginia.
What it is
According to Alibaba's announcement and the Model Studio model page:
- Size: a mixture-of-experts model (only part of the network runs for each word) with 2.4 trillion parameters in total and about 71 billion active per token. Alibaba says that is 25% fewer active parameters than Qwen3.8-Max.
- Inputs and outputs: text, images and video in; text out.
- : 1 (a token is roughly three-quarters of a word). Answers can run to 131,072 tokens, or 262,144 tokens including the model's thinking.
- Reasoning: thinking is on by default, with three effort levels (low, high and xhigh).
- Tools: , parallel tool calls, web search and structured JSON output.
- Compatibility: the API accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, as well as Alibaba's own DashScope format. In practice, many apps built for OpenAI or Claude can try it by changing the endpoint, key and model ID (
qwen4-max). - Not supported: batch processing, fine-tuning and private deployment, per the model page.
What it costs
Alibaba's pricing page (updated 16 November) lists one price band for prompts up to 1 million tokens, the same for thinking and non-thinking modes, in US dollars per million tokens:
Quick check · Everyday ↔ real
Match the colleague to the AI
Switch between the two and see how each part lines up.
- Reading a letter
- Looking at a photo
- Listening to a voicemail
- Jotting it all on one notepad
- Writing back one reply
A colleague at the next desk who can read, look and listen
| Everyday | Real |
|---|---|
| Reading a letter | Text input |
| Looking at a photo | Image input |
| Listening to a voicemail | Audio input |
| Jotting it all on one notepad | Turning inputs into tokens |
| Writing back one reply | The model's answer |
From the explainer What is multimodal AI? Explained with a colleague who can read, look and listen
| Region | Input | Output | Cache read |
|---|---|---|---|
| Singapore (International) | $1.50 | $4.50 | $0.075 |
| Frankfurt, Virginia, Tokyo, Hong Kong (Global) | $1.413 | $4.24 | $0.071 |
| Beijing (Chinese mainland) | $1.413 | $4.24 | $0.071 |
Output includes the model's thinking, so long reasoning runs cost more than the visible answer suggests. The Singapore region comes with a free quota of 1 million tokens, valid for 90 days; the page lists no free quota for the other regions.
How Alibaba says it performs
All the performance figures come from Alibaba itself. It says Qwen4-Max ranks sixth on the Code Arena coding leaderboard. On Work360, an internal test Alibaba built covering 721 specialist fields, it says the model beats Qwen3.8-Max while using 35% fewer tokens at high effort and cutting the average cost per task by around 60%. The release also names several other benchmarks without giving scores. Treat these as the vendor's claims until independent results and your own tests agree.
Limits and caveats
- Not open weights. The announcement does not say Qwen4-Max itself will be downloadable. It says smaller open-weight Qwen4 models will be released soon.
- The preview is ending. If you tested
qwen4-max-preview, the preview page says it runs until midnight on 25 November 2026 (UTC+8) and will then be retired. Switch toqwen4-max. - Region and data rules. Model Studio has separate deployment scopes (Chinese mainland, International and Global). Check which region your account and API key use, and what your contracts or data-protection rules say about the provider, before you send customer data.
Who should care
Teams running coding assistants or multi-step agents on another provider's API may want a lower-cost option to benchmark, especially since the OpenAI- and Anthropic-compatible endpoints keep the switching work small. For comparison, check your current provider's per-million-token prices against the table above using your own typical prompts, including thinking tokens.
What to do this week
- Pick five to ten real tasks from your product (a bug fix, a document summary, a multi-tool workflow) and run them on Qwen4-Max and your current model, comparing quality, tokens used and cost.
- Move any tests still on
qwen4-max-previewbefore 25 November. - Ask whoever owns compliance whether the Frankfurt or Virginia regions meet your data requirements before using it in production.
- Keep a fallback model in your routing setup so one provider's outage or price change doesn't stop your app.
The DevDrift helps teams test and switch AI models inside their products, with approvals and audit trails kept in place.
Sources
You’ve read the news. Now understand it.
- Learn · 5 min · Everyday ↔ realWhat is multimodal AI? Explained with a colleague who can read, look and listenMultimodal AI is an AI model that can take in more than one kind of information, such as text, pictures and sound, and use them together. Think of a colleague who can read, look and listen.Start in 5 seconds →
- Learn · 5 min · Try the dialWhat is a context window? Explained with a clerk's deskA context window is how much text an AI model can look at in one go: your instructions, documents, the chat so far and its own reply. Explained with a clerk working at a desk of fixed size.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat are tokens in AI? Explained with word-game tilesTokens are the small pieces of text an AI model reads and writes, and the unit used to measure limits and cost. Explained with a word game whose tiles carry chunks of words.Start in 5 seconds →
- Learn · 5 min · Put it in orderWhat is tool use in AI? Explained with a calculator in an examTool use, or function calling, lets an AI model ask your software to run a tool, like a lookup or a calculator, and use the result. Explained with a student allowed a calculator in an exam.Start in 5 seconds →


