Alibaba has launched Qwen4-Max, the flagship model of its new Qwen4 family — and priced it to undercut its American rivals. The Qwen4-Max price is $1.50 per million input tokens and $4.50 per million output tokens through Alibaba’s QwenCloud service, with cached input at $0.075 per million. On Alibaba’s own Work360 benchmark of industry tasks, the average cost of completing a task fell from $4.43 with the previous model to $1.67 — a 62% cut. Announced on Monday, November 16, it is the first release in the Qwen4 generation and the successor to Qwen3.8-Max. See Qwen’s official announcement for details.
What Qwen4-Max is
Qwen4-Max is built on a new architecture with a 2.4-trillion-parameter backbone, of which 71 billion parameters are active for each token, plus 328 billion N-gram embedding parameters. It has native vision — it can process images alongside text — and a one-million-token context window, long enough to hold roughly a thousand pages of documents in a single conversation. Alibaba describes it as a “24/7 AI teammate” for professional work: 3D modelling in Blender, building games in Unity and Unreal Engine, financial research, legal due diligence and clinical record review.
The announcement leans heavily on long-running agent work. In one showcase, the model ran a 48-hour autonomous session that designed a new coolant molecule for data-centre chips, screening 2.7 million candidates and directing an automated laboratory to synthesise and test a physical sample. In another, it made 454 tool calls across a drug-discovery workflow targeting Alzheimer’s-related inflammation, producing 24 candidate molecules. These are Alibaba’s own demonstrations, not independent results.
How to access it
Qwen4-Max is available now on QwenCloud, either through a Token Plan subscription or pay-as-you-go API access, under the model ID qwen4-max. It is also available in Alibaba’s QwenWork, Qoder and QwenStudio products. A Token Plan-only preview, qwen4-max-preview, opened on November 11 and runs until November 25, when it will be retired; the production model is the one to use.
One break with Qwen tradition matters: unlike earlier Qwen flagships, Qwen4-Max is not open-weight. Its weights cannot be downloaded and run on private hardware. Alibaba says smaller Qwen4 models will be released as open source later, but has given no date — and nothing in the announcement promises Max itself will follow.
The independent verdict is cooler
Alibaba’s own comparison chart puts Qwen4-Max ahead of Claude Opus 5.5 and GPT-6 Astra on one coding test, DeepSWE, and well ahead on a legal agent test. But the first independent score tells a different story. Artificial Analysis, which runs nine evaluations through the same harness for every model, scores Qwen4-Max at 43 on its Intelligence Index — 38th of 81 models, just above the median of 41. On the same index, Claude Opus 5.5 scores 59, GPT-6 Astra 57 and Moonshot’s Kimi K3.1 Pro 46. Qwen4-Max also trails badly on Terminal-Bench 4.0, a terminal-coding test, at 48.5 against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1 Pro.
The price is where the model genuinely competes. At $3.27 per Intelligence Index task, it costs about a third of Claude Opus 5.5’s $10.04 — and only OpenAI’s GPT-6.1 Sol is cheaper, at $1.27. On LM Arena’s web-development leaderboard it ranks sixth, behind Claude Fable 5.2, Opus 5.5, Gemini 4 Carbon and GPT-6 Astra. The honest summary: a capable, cheap workhorse for professional agent tasks, not a new frontier leader.
Safety claims, in Alibaba’s words
Alibaba says it strengthened the model’s defences: on its secure-coding tests, the share of code generated both correctly and securely rose from 8% with Qwen3.8-Max to 14.7%, and internal reviewers found no capabilities posing large-scale cybersecurity risk. It also says it added safety data for chemical, biological, radiological and nuclear topics during post-training. All of these assessments are Alibaba’s own; no external red team has published results yet.
Why it matters beyond China
Qwen4-Max arrives in a week crowded with flagship releases — Google’s Gemini 4 Hydrogen and Anthropic’s Claude Fable 5.2 both launched within the past few days. Alibaba’s answer is not to outscore them but to undercut them: a model good enough for long professional jobs at a third of the cost per task. If developers accept that trade, the pricing pressure will reach every lab’s API price list. The open question is whether the closed-weights decision costs Alibaba the developer goodwill that made Qwen the world’s most-downloaded open-weight family.
What to watch
Three things will decide how far Qwen4-Max travels. First, independent benchmarks beyond Artificial Analysis — the vendor’s own charts flatter it, and the gap on terminal coding is large. Second, whether the promised open-weight Qwen4 models appear, and when. And third, how rivals answer on price.
