
The closing card of Alibaba’s Qwen4 launch video. Image: Alibaba (Qwen)
Alibaba has released Qwen4-Max, its new flagship AI model and China’s biggest answer yet to GPT and Claude: a 2.4 trillion-parameter model with a 1 million-token context window that, Alibaba says, finishes the same work for 62% less than the model it replaces. Its own benchmark charts tell a more mixed story, with Anthropic’s Claude Opus 5.5 still ahead on most coding tests.
What Qwen4-Max is
Qwen4-Max is the first model in Alibaba’s Qwen4 family and takes over from Qwen3.8-Max. It reads text, images and video, replies in text, and reasons before it answers by default. It is a mixture-of-experts model: 2.4 trillion parameters in total, of which 71 billion do the work for each token, plus 328 billion extra “N-gram embedding” parameters. Alibaba says that is 25% fewer active parameters than Qwen3.8-Max, and that training ran with “no loss spikes”, the sudden failures that often dog runs this size.
It is no bigger than Qwen3.8-Max, which was also 2.4 trillion parameters, even though Alibaba said in September it plans to train a 5 to 10 trillion-parameter model. Alibaba pitches Qwen4-Max as a “24/7 AI teammate” for long jobs: building games in Unity, Unreal Engine and Godot, 3D modelling in Blender, finance, law and healthcare work, and a lab project in which it says Qwen4 designed and made a new coolant material. Those are company demos, not independent tests.
Cheaper, by Alibaba’s numbers
The headline claim comes from Work360, Alibaba’s own benchmark covering 721 kinds of professional work. There, the average cost of a task fell from $4.43 with Qwen3.8-Max to $1.67, a 62% cut, because the new model uses fewer tokens (about 35% fewer at high effort) and costs less per token. On Terminal-Bench 4.0, a coding test, it used about 49% fewer tokens.
On QwenCloud, Alibaba’s international developer platform, the API costs $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075. OpenRouter lists Qwen3.8-Max (0902) at $2 and $6, and Claude Opus 5.5 lists at $4 and $20, so Qwen4-Max’s output costs less than a quarter of Opus’s.
Where it wins, and where it doesn’t
Alibaba compares Qwen4-Max with Claude Opus 5.5, OpenAI’s GPT-6.1 Sol and GPT-6 Astra, and Moonshot’s Kimi K3.1 across 16 tests. These are the company’s own figures:
- Ahead: DeepSWE v1.1, an agentic coding test, by a whisker (75.7 against 75.5 for Kimi and 74.2 for Opus); Agents’ Last Exam (46.7 against 43.4 for Opus and Kimi); and HLAB-Vals legal agent work (21.3 against 9.2 for Opus).
- Behind: Terminal-Bench 4.0 (48.5 against 59.6 for Opus and 67.2 for Kimi); ProgramBench (31.5 against 65.0 for Opus); JobBench (68.5 against 75.4); and Humanity’s Last Exam (51.7 against 59.8).
The independent picture is cooler still. Artificial Analysis gives Qwen4-Max 43 on its Intelligence Index, up from 34 for Qwen3.8-Max (0902), but behind Claude Fable 5.2 (65), Claude Opus 5.5 (59), GPT-6 Astra (57) and Moonshot’s Kimi K3.1 Pro (46). On that measure it is not even China’s top-scoring model.
Tighter cyber guardrails
Alibaba says it filtered Qwen4-Max’s training data by purpose, keeping material for defensive security work such as finding and fixing vulnerabilities and removing data meant to boost attacks. It also checked training code for security as well as function: the share of its code that both works and is secure rose from 22.0% to 31.8% on one test, Alibaba says.
Its internal security experts reviewed long task runs, including the model’s full chain of thought, and wrote: “No capabilities posing large-scale cybersecurity risks have been observed.” On chemical, biological and nuclear weapons, they found no risk of “significant uplift”. Both are Alibaba’s own reviews; no outside safety tests were published, a gap that is common among Chinese AI labs. Last month Cohere said an open Qwen model tried to upload a file to Beijing in a security test.
Where you can use it
Qwen4-Max is closed: Alibaba has not released its weights, though it says smaller Qwen4 models “are coming to open source”, with no date. Developers can reach it through QwenCloud’s API or its Token Plan subscription, priced in US dollars, and through OpenRouter as qwen/qwen4-max. Everyone else can try it in Alibaba’s Qwen Studio chat at chat.qwen.ai, its QwenWork assistant and the Qoder coding tool.
Why it matters
Qwen4-Max narrows the gap with US labs and cuts the price sharply, but on Alibaba’s own charts and the independent index it still sits a tier below the best American models, and below a Chinese rival. For companies whose AI bill is driven by long, token-hungry agent jobs, a task that costs 62% less may matter more than a few benchmark points.
Sources: Qwen (Alibaba), QwenCloud, Artificial Analysis, OpenRouter.


