Companies already building on Qwen3.8-Max can move to a stronger flagship at lower list prices in China and abroad, though Qwen4-Max offers no downloadable weights.
Alibaba has made Qwen4-Max, the first model of its Qwen4 generation, generally available to developers, and it is charging less for it than for the model it succeeds. The release on November 16 moved the model onto Alibaba’s QwenCloud API and into its Qwen Studio, QwenWork and Qoder apps, The Next Web reported, with list prices of US$1.50 per million input tokens and US$4.50 per million output tokens.
The launch follows the 15-day, subscription-only preview that Alibaba opened on November 11, when the model could be used only through its Token Plan subscriptions, with no per-token price and no benchmark results. Both gaps are now filled. Alibaba describes Qwen4-Max as a mixture-of-experts model with 2.4 trillion total parameters, native vision and a one-million-token context window, and has published its own performance claims alongside the prices.
Those claims are Alibaba’s. According to PYMNTS, Alibaba said Qwen4-Max outperformed Anthropic’s Claude Fable 5.1 and Moonshot AI’s Kimi K3.1 Pro, without publishing scores for those comparisons. The Next Web reported the same comparison, according to Bloomberg, and said it had not verified Alibaba’s scores; Alibaba’s English blog post does not repeat it.
The general release adds prices, a model size and Alibaba’s benchmark set
Alibaba Cloud’s launch post, datelined Hangzhou, says Qwen4-Max activates about 71 billion of its 2.4 trillion parameters for each token, 25% fewer than Qwen3.8-Max at the same total size. A mixture-of-experts model routes each request through only part of its network, which cuts the computing cost of every answer. Alibaba calls Qwen4-Max the most capable Qwen model to date.
The model ranked sixth on Code Arena, a public coding leaderboard, and Alibaba reported strong results on agent and workplace tests including Agents’ Last Exam, Vals Index and JobBench. Its headline figure comes from Work360, Alibaba’s own test of industry tasks across 721 specialized fields. There, Alibaba says, Qwen4-Max beat Qwen3.8-Max while using 35% fewer tokens at its high reasoning setting, and cut the average cost per task by about 60%. TechNode reported the change as a fall from RMB26.58 to RMB11.16 per task.
Alibaba’s developer documentation confirms the one-million-token window and shows that the model can return up to 131,072 tokens in a reply without reasoning, or 262,144 tokens of reasoning and answer combined. Reasoning is on by default at the high setting, and the model reads both images and video.
Independent results show a narrower lead than Alibaba’s charts
Outside testing is starting to arrive. On the Vals Index, which grades models on professional tasks, Vals AI scores Qwen4-Max at 63.20%, 11th of 38 models, just ahead of Kimi K3.1 Pro at 62.52%. The Vals listing also shows a newer Anthropic model, Claude Fable 5.2, dated November 9, which was not in Alibaba’s comparison.
Alibaba’s own charts are mixed in places. The Next Web, citing FourWeekMBA’s reading of them, said Qwen4-Max scored 75.7 on the DeepSWE v1.1 coding test against 74.2 for Claude Opus 5.5, but 48.5 on Terminal-Bench 4.0, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2.
Price is the clearer advantage, and Alibaba Cloud’s pricing page, updated on November 17, sets out the cut. Qwen4-Max lists at US$1.50 per million input tokens and US$4.50 per million output tokens on Alibaba’s Singapore site, down from US$2 and US$6 for Qwen3.8-Max. In mainland China, its Chinese pricing page lists RMB10 for input and RMB30 for output per million tokens, against RMB12 and RMB36 for Qwen3.8-Max, with cache hits at RMB0.5.
Open-weight Qwen4 models are the next release to watch
Qwen4-Max itself stays closed: developers cannot download or modify it. Alibaba says smaller open-weight models from the Qwen4 series "will soon be released to developers worldwide," without giving dates or sizes. Those releases matter most for the companies and labs that build on Qwen’s open models rather than renting Alibaba’s cloud.
For the closed flagship, the test now moves to paying users. Further independent leaderboards will show whether the narrow Vals lead over Kimi K3.1 Pro holds, or whether the weaker Terminal-Bench numbers in Alibaba’s own charts are the better guide to how Qwen4-Max handles long agent work.