Alibaba opened its Qwen4-Max model to pay-as-you-go API customers on 16 November at $1.50 per million input tokens and $4.50 per million output tokens. The Model Studio release notes list qwen4-max for the international deployment that day. The weights stay closed, and the first outside scores put it in the upper middle of the pack, not at the top.

A preview went out on 11 November, but only through Token Plan. Per the preview documentation, that model is retired after 25 November.

What a developer gets

The QwenCloud model page lists a 1M-token context, up to 991K tokens of input, up to 393K tokens of output and a reasoning budget of up to 262K tokens. It takes text, images and video and returns text.

The API is compatible with the OpenAI and Anthropic formats, and web search, function calling and structured JSON output are built in.

The price, against its own predecessor

Implicit cache reads, meaning repeated prompt prefixes the service has already processed, cost $0.075 per million tokens. The Qwen3.8-Max page lists $2 input, $6 output and $0.25 for cache reads. That makes the new model 25% cheaper per input and output token and 70% cheaper on cached input.

With made-up numbers, take a coding agent session that reads 1M tokens, 800K of them cached, and writes 200K. At list prices that costs about $1.80 on Qwen3.8-Max and $1.26 on Qwen4-Max.

What Alibaba measured, and who ran it

Alibaba’s launch post describes a mixture-of-experts model: a router sends each token to a few expert sub-networks, so only 71 billion of its 2.4 trillion parameters work on any one token.

The headline efficiency claim comes from Work360, a benchmark Alibaba built itself, covering tasks in 721 specialized fields. On it, Alibaba says Qwen4-Max beats Qwen3.8-Max while using 35% fewer tokens at high reasoning effort and cutting average cost per task by around 60%. The post names no grading method or outside party that reran it.

Other claims are demonstrations, such as cutting manual workpiece programming from roughly 475 minutes to 24, with no detail on the part or how the result was checked.

What outside testers found on day one

Artificial Analysis ran its own Intelligence Index, nine evaluations covering coding, agent work and knowledge. Qwen4-Max scored 43, against a median of 41 for models in its price tier, at an average cost of $3.27 per task. It generated 127M output tokens on that run, against a median of 78.6M.

A lower per-token rate helps less if the model writes more tokens to get there.

On Arena’s Code Arena WebDev leaderboard, which ranks models by user votes on web-development tasks, Qwen4-Max sits sixth with a preliminary score of 1,750 from 1,814 votes. Its listed rank spread runs from 3 to 9. We explained what such vote-based boards do and do not measure in our guide to the four big AI leaderboards.

Not open weights

Alibaba says smaller open-weight Qwen4 models will follow “soon”, without sizes or dates. Earlier open Qwen releases became bases for outside work, such as Nous Research’s Arete-27B retune of Qwen3.8-27B. If you need local inference, see our explainer on small language models.

What to do this week

If you call Qwen3.8-Max or the preview, rerun your own evaluation prompts on qwen4-max at the same reasoning effort and log output tokens and cost per task. The preview documentation says that, starting with qwen4-max-preview, max_tokens caps thinking and the final answer together, so check limits sized for answers alone. Move preview traffic before 25 November.

Sources

Lars Strother

Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.