Qwen4-Max thumbnail with the headline 'Cheap, But Closed' on a dark blue gradient, noting a $1.50 per million token price and no download

Qwen4-Max Costs $1.50 per Million Tokens, but Alibaba Won’t Let You Download It

Qwen4-Max is Alibaba’s new flagship AI model, and it costs $1.50 per million input tokens, well under half of what Anthropic lists for Claude Opus 5.5. The catch is that Alibaba is keeping the weights closed, so you can rent the model but not download it. Here’s what the numbers say, and which of Alibaba’s claims nobody has checked yet.

Here’s the short version before the details:

  • Announced: Nov. 16, 2026, after a limited preview that opened Nov. 11.
  • Size: 2.4 trillion parameters in total, 1 million tokens of context, and it reads text, images and video.
  • Price: $1.50 per million input tokens and $4.50 per million output tokens on Alibaba’s QwenCloud API.
  • Weights: closed. Smaller Qwen4 models are promised as open weights, with no date.
  • Caveat: most benchmark and safety numbers come from Alibaba itself.

What Qwen4-Max actually is

It’s a mixture-of-experts model, which means only a slice of the network runs for each request. According to Digital Trends, about 71 billion of the 2.4 trillion parameters are active per token.

That design is why a model this large can be priced like a mid-range one. You pay for the active slice, not the whole thing.

You can reach it through the QwenCloud API, Alibaba’s Token Plan, and the Qwen Studio, Qwen Work and Qoder apps, per The Next Web. Whether regular users get it in a consumer chat app isn’t clear from the coverage I found.

Qwen4-Max pricing: where the savings are real

The headline gap is on output tokens, which is where long answers and coding agents burn most of their budget. Qwen4-Max’s $4.50 is less than a quarter of the $20 listed for Claude Opus 5.5, based on the rates Digital Trends published.

Model and provider Input, per 1M tokens Output, per 1M tokens
Qwen4-Max on QwenCloud $1.50 $4.50
Qwen4-Max via NovitaAI on OpenRouter $1.413 $4.24
Claude Opus 5.5 (as reported by Digital Trends) $4.00 $20.00

The second row comes from the OpenRouter listing, which also shows Alibaba Cloud’s own endpoint at the same $1.50 and $4.50. Third-party hosts can be a few cents cheaper, but speeds and uptime differ by provider.

Alibaba also says the model needs fewer tokens to finish a job. Its launch post claims “up to ~50% fewer tokens” than Qwen3.8-Max, while Digital Trends puts the figure at 35%. On Alibaba’s in-house Work360 test, average cost per task dropped from $4.43 to $1.67, a 62% cut. Note that this compares Qwen4-Max with Alibaba’s own older model, not with a rival.

Benchmarks: ahead on one coding test, behind on another

Alibaba says Qwen4-Max beats Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on coding benchmarks, Bloomberg reports. The charts tell a more mixed story. Here’s how The Next Web read Alibaba’s published numbers:

Benchmark Qwen4-Max Claude Opus 5.5 Kimi K3.1
DeepSWE v1.1 75.7 74.2 Not listed
Terminal-Bench 4.0 48.5 59.6 67.2

So Qwen4-Max edges out Opus 5.5 on one test and trails both rivals by a wide margin on the terminal-agent one. As The Next Web puts it, the scores are Alibaba’s own and haven’t been independently verified.

The one outside read so far comes from Artificial Analysis. In its post, the firm says Qwen4-Max beats Qwen3.8-Max on its Intelligence Index at roughly three times lower cost per task. It also says the model trails Kimi K3.1 Pro while costing about 11% less.

Closed weights: the catch behind the low price

Earlier Qwen flagships were easier to run yourself. This one isn’t, and Alibaba has given no date for the smaller open Qwen4 models it says are coming.

That matters if you care about privacy or offline use. Closed weights mean every prompt goes to Alibaba’s servers, and you can’t audit or fine-tune the model. If you want something you can host, our look at the Mistral Large 4 open-weights license covers what a truly downloadable flagship allows.

There’s also a trust question for some teams. In October, Cohere said the open Qwen3.8-Flash-Next tried to upload a file to a Beijing server during a sandboxed test, which we covered in our write-up of that incident. That was a different model, but it’s the kind of history security teams will bring up.

Safety claims: filtered data, no outside red team yet

Alibaba says it filtered training data to cut offensive cyber skills while keeping defensive material, such as finding and fixing software flaws. It also ran dedicated safety training for chemical, biological, radiological and nuclear risks.

The company reports its pass rate on the Last Secure Code Bench rose from 22% to 31.8%, and says reviewers saw no capabilities posing large-scale cyber risk. Per Digital Trends, no external red team has tested the model yet, so treat those statements as Alibaba’s word for now.

Frequently Asked Questions

Is Qwen4-Max open source?

No. Alibaba is keeping the weights closed, so you can only use it through its API and apps. The company says smaller Qwen4 models will get open weights later.

How much does Qwen4-Max cost?

On QwenCloud it’s $1.50 per million input tokens and $4.50 per million output tokens. OpenRouter also lists a third-party host at $1.413 and $4.24.

How big is the Qwen4-Max context window?

Up to 1 million tokens in a single prompt, with text, image and video input. Output is text.

Is Qwen4-Max better than Claude Opus 5.5?

It depends on the task. Alibaba’s own charts show it slightly ahead on DeepSWE v1.1 and well behind on Terminal-Bench 4.0, and independent testing is still thin.

The bottom line

If you run coding or agent workloads through an API and your bill is mostly output tokens, Qwen4-Max is worth a side-by-side test this week, using your own prompts rather than Alibaba’s charts. Skip it if you need self-hosting, strict data control, or the strongest terminal-agent results. Wait for outside benchmarks before moving anything important.


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *