Qwen3.8 Max Test: 13 Tasks, Scores, Pricing & Access

👁 123 reads

Qwen3.8 Max writes correctly and follows instructions to the letter, but thinks slowly and you pay for that thinking. In our lab, Alibaba’s flagship model (version 0902) scored 76.4 out of 100 in writing, met all 26 rules from prompts, and flawlessly solved difficult proofreading and a quiz. Below you’ll find results task by task, response excerpts, prices in dollars, and one correction we need to make to our own ranking.

What This Model Is

Qwen3.8 Max is Alibaba’s largest model. The 0902 snapshot was released on September 2, 2026. According to the OpenRouter description, it’s a mixture-of-experts model with 2.4 trillion parameters that accepts text, image, and video, and responds with text. Alibaba Cloud documentation lists a 1 million token context and up to 131,072 output tokens. The weights are closed. Reasoning is enabled by default; in Alibaba’s API you can disable it with the enable_thinking parameter or limit it through thinking_budget.

In the Artificial Analysis index (September 27, 2026) it has 45 points, the same as GLM-5.3 in the max variant. It generates 39 tokens per second and is the most expensive among top models per task: $5.41 USD, because in the index it produced 190 million tokens. We described the entire model family in our Qwen AI guide.

Our Results: Task by Task

The test is the same as in our writing ranking: 13 prompts through OpenRouter, effort=low, blind evaluation by three judges from other companies (Claude Opus 5.5, GPT-6 Sol, Gemini 3.1 Pro). We ran the writing tasks in Polish, a heavily inflected language, so this is a non-English writing test - a useful proxy for multilingual performance. The scores below come from the September 27 round, in which Qwen was an anchor alongside new models. Its result differed from the first round by 0.5 points (75.9 versus 76.4).

Average judge scores on a 1-10 scale. * Second attempt with higher limit, explanation below.

Qwen passed the remaining tests almost flawlessly: in easy proofreading 21 out of 22 corrections, in difficult proofreading 30 out of 30, in quizzes 46 out of 46 and 48 out of 48. Claude Opus 5.5 had the same perfect scores in difficult proofreading and the difficult quiz.

Response Excerpts

The magical realism short story (8.56) begins like this (translated from the model’s Polish output):

“Halina had worked at the post office for thirty-two years. She knew every block in the Millennium housing estate, every stairwell, every smell - from boiled cabbage on the ground floor to mothballs on the fourth floor.”

The OpenAI judge found it coherent with well-built tension, and Opus 5.5 noted the strong, open ending. Two judges had reservations about length, though our counter showed 305 words, within the prompt’s range (300-350).

In the newsletter translation, idioms came out naturally, without literal calques:

“Let’s be honest: onboarding new clients isn’t rocket science, but we’ve been cutting corners for too long. (…) It was a bucket of cold water. Losing those clients turned out to be a blessing in disguise.”

Here too you can see a weakness: “onboardowanie” and “follow-up” remained in English, and one phrase changed meaning (clients left, rather than were not acquired).

The weakest summary (6.22) had an error that all three judges caught: a grammatical mistake (in the original Polish output). Qwen also omitted training and the forecasting module. GLM-5.3 made the same error.

Thinking Costs More Than Text

With effort=low, Qwen generated 37.7k tokens across 13 tasks, of which 33k were thinking. The most expensive was the short product description:

• September 26, first attempt: 6,000 thinking tokens, exhausted limit, empty response, $0.0366 USD and almost 6 minutes.
• Second attempt with higher limit: 14,460 thinking tokens, $0.0897 USD, almost 5 minutes. The same description took GPT-6 Luna 3.6 seconds.

Judges evaluated this second attempt, and that was an error in our method: other models didn’t get a higher limit. If you count the product description as failed, Qwen has 68.2 points instead of 76.4, roughly the same as MiMo-V2.6-Pro (69.1) and clearly less than Muse Spark 1.3 (75.5). We report both numbers in our comparison of lesser-known top models. With higher reasoning settings this problem grows: in our Chinese models test, the Prime version counted every word in its thoughts and didn’t return text.

Pricing and Our Test Cost

Alibaba charges one rate across the entire context range, without tiers. We verified prices in the Model Studio price list (international version, Singapore region) and on OpenRouter’s list, where Alibaba is the only provider.

How to Access Qwen3.8 Max

• OpenRouter - model qwen/qwen3.8-max-0902, same price as Alibaba. This is how we ran our test.
• Alibaba Cloud Model Studio - international version, account with card. In the native API you have enable_thinking and thinking_budget parameters.
• Qwen Chat - Alibaba’s free chat. We couldn’t confirm whether version 0902 is available there.

For writing tasks, we’d set low reasoning or a thinking budget, because otherwise you pay for thousands of tokens you won’t see. That’s our opinion, based on the numbers from this test.

Frequently Asked Questions

Does Qwen3.8 Max write well?

Correctly: 8.81 out of 10 for linguistic accuracy in judges’ evaluation. Worse for naturalness (7.11) - texts can be formulaic. Combined score 76.4 points, with equal limits 68.2.

How much does Qwen3.8 Max cost?

$2 USD per million input tokens and $6 USD per million output tokens, excluding VAT. Our 13-task test cost $0.24 USD.

Why is Qwen3.8 Max so slow?

Because it thinks a lot and generates around 39 tokens per second. In our test, 9 short texts took it 572 seconds, GPT-6 Luna 39.

Is Qwen3.8 Max open source?

No. The Max version weights are closed; smaller models in the family are open, such as Qwen3.8 27B.

Kinetyra