Alibaba Cloud made Qwen4-Max generally available on its Model Studio platform on November 16, 2026, five days after a preview limited to subscribers. The first model of the Qwen4 generation reads text, images and video, reasons by default, holds up to 1 million tokens of context and costs $1.50 per million input tokens and $4.50 per million output tokens in Alibaba’s Singapore region, 25% below the Qwen3.8-Max list price. It is a closed, API-only model: there are no downloadable weights and no fine-tuning.

What Alibaba released

Alibaba’s documentation describes Qwen4-Max as the next flagship of the Qwen family, built for coding, office work, front-end design and long agentic tasks. The model takes text, image and video input and answers in text. It supports streaming, web search, parallel tool calling, function calling, prefix completion and structured JSON output.

Office buildings of Alibaba's Xixi Park campus in Hangzhou seen from outside
Alibaba's Xixi Park in Hangzhou, home of the group whose cloud unit develops the Qwen models. Photo: Danielinblue / Wikimedia Commons, CC BY-SA 4.0

Developers can reach it through four interfaces: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Alibaba’s own DashScope API. That compatibility matters in practice, because most coding agents and agent frameworks already speak one of the first three, so switching a backend can be a configuration change rather than a rewrite. Vercel added the model to its AI Gateway on November 15 under the name alibaba/qwen4-max.

The model is listed in six regions: Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. Batch inference is available only in Beijing, and the documentation marks fine-tuning and custom deployment as unsupported everywhere.

How reasoning and limits work

Thinking is switched on by default at a “high” effort level, with “low” and “xhigh” as the other options. The documentation maps other vendors’ effort names onto these three, so a client that sends “medium” gets “high” and one that sends “max” gets “xhigh”.

The numbers that matter for long jobs:

  • Context window: 1,000,000 tokens.
  • Default thinking budget: 131,072 tokens.
  • Thinking plus final answer: at most 262,144 tokens combined.
  • Maximum non-thinking output: 131,072 tokens.

Because reasoning and the visible answer share one output allowance, a model that thinks for a long time leaves less room for the response, and every thinking token is billed as output.

How Qwen4-Max pricing compares

In Singapore, Qwen4-Max lists at $1.50 per million input tokens and $4.50 per million output tokens, against $2 and $6 for Qwen3.8-Max. Cached input costs $0.075 per million tokens; writing an explicit cache costs $1.875 per million tokens for five minutes or $3 for one hour, and the cached prefix must be at least 1,024 tokens. Beijing and Hong Kong list slightly lower prices ($1.413 and $4.24). New Singapore accounts get a free quota of 1 million tokens for 90 days.

Here is how the list prices stack up against the other frontier models in our directory, using the prices recorded on each model page:

Model Input ($/1M tokens) Output ($/1M tokens) Context
Qwen4-Max (Singapore) 1.50 4.50 1M
Gemini 4 Hydrogen 1.50 7.50 1M
GPT-6.1 Sol 2.00 10.00 1.05M
Claude Opus 5.5 4.00 20.00 1M

Qwen4-Max matches Gemini 4 Hydrogen on input and is the cheapest of the four on output. Output price matters most for reasoning models, because hidden thinking tokens are billed as output. List prices only tell part of the story, though. Alibaba has published no independent leaderboard results for Qwen4-Max yet, and a cheaper model that needs more reasoning tokens or more retries can end up costing more per finished task.

Preview, subscription and who can use it

The release follows Qwen4-Max-Preview, which opened on November 11 (UTC+8) only through Alibaba’s Token Plan subscription. The preview runs until November 25, after which it will be retired. The full model is also included in the Singapore Personal Edition Token Plan, which bills in credits and works with interactive coding tools such as Claude Code, Cursor and Qwen Code. The plan’s terms bar automation scripts, custom application backends and non-interactive batch calls, so production apps need ordinary pay-as-you-go API access.

According to Vercel’s announcement, Alibaba says Qwen4-Max uses fewer tokens than earlier Qwen Max models to finish the same tasks, writes code with fewer injection and memory-safety flaws, and can drive professional tools such as Blender and game engines. These are vendor claims. We have not seen independent measurements of them.

What it means for developers

Teams already using the Qwen3.8 open-weight family for local work now have a hosted flagship that speaks the same APIs as their other providers. The practical trade-offs are clear: lower list prices and a long context window, against no weights, no fine-tuning and batch processing only in Beijing. For agents that resend long system prompts and tool definitions, the caching rules (a minimum prefix, and five-minute or one-hour lifetimes) will shape the real bill as much as the headline rate.

What to watch next

  • Independent leaderboard results for Qwen4-Max, which will show whether the lower price comes with frontier-level quality.
  • Whether Alibaba releases open-weight Qwen4 models, as it did for earlier generations.
  • The retirement of Qwen4-Max-Preview on November 25 and any changes to Token Plan terms.
  • Price responses from Google, OpenAI and Anthropic on output tokens for reasoning models.

Key takeaways

  • Qwen4-Max is generally available on Alibaba Cloud Model Studio as of November 16, 2026.
  • It accepts text, image and video, reasons by default and offers a 1 million-token context window.
  • Singapore list prices are $1.50 and $4.50 per million input and output tokens, 25% below Qwen3.8-Max.
  • It works through the OpenAI- and Anthropic-compatible APIs, but has no open weights and no fine-tuning.
  • Alibaba’s quality claims have not been confirmed by independent leaderboards yet.

Frequently asked questions

Is Qwen4-Max open source?

No. Qwen4-Max is available only as a hosted API on Alibaba Cloud Model Studio and through partners such as Vercel’s AI Gateway. Alibaba has not released its weights, and fine-tuning is not supported.

How much does Qwen4-Max cost?

In Alibaba’s Singapore region it lists at $1.50 per million input tokens and $4.50 per million output tokens. Cached input costs $0.075 per million tokens, and new Singapore accounts get 1 million free tokens for 90 days.

Can I use Qwen4-Max with Claude Code or OpenAI SDKs?

Yes. The model supports the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages interfaces, so tools built on those APIs can call it. Alibaba’s Token Plan subscription also lists Claude Code, Cursor and Qwen Code as supported interactive tools.