Skip to content
Products

Alibaba Releases Qwen4-Max, Cutting Output Price 25% to $4.50 per Million Tokens

The closed-weight flagship claims about 62 per cent lower average task cost on Alibaba's own benchmark, but trails Kimi K3.1 and Claude Opus 5.5 on Terminal-Bench 4.0.

·2 min read

Prefer Nakoda News on Google

Alibaba made Qwen4-Max, its new closed-weight flagship AI model, generally available on Monday, China time, at $1.50 per million input tokens and $4.50 per million output tokens, 25 per cent below Qwen3.8-Max. It is offered through the QwenCloud API and Alibaba's Qwen Studio, QwenWork and Qoder products, after a preview that opened to Token Plan subscribers on November 11. Details were first published by Bloomberg.

The model became generally available on Monday, China time, after a preview that opened on November 11 to subscribers of Alibaba's Token Plan. It is offered through the QwenCloud API and Alibaba's Qwen Studio, QwenWork and Qoder products. Unlike many earlier Qwen releases, it is not open-weight: Alibaba has said smaller Qwen4 models will be released as open source later, without giving sizes or a date.

Price is the main pitch. On Alibaba's international Model Studio list, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, 25 per cent below the $2 and $6 charged for Qwen3.8-Max. Cached input reads are priced at $0.075 per million tokens. Alibaba says the model reaches answers using fewer tokens, and on its internal Work360 benchmark the average cost per task fell to $1.67 from $4.43, a reduction of about 62 per cent. One outlet, using yuan figures, put the cut at 58 per cent.

The model takes text, images and video as input and has a context window of 1 million tokens. One report on the preview said the base context was 262,000 tokens, extendable to the full million, so very long inputs may behave differently from the headline figure. Technical write-ups describe a mixture-of-experts design with 2.4 trillion total parameters and 71 billion active per token, though at least one outlet said Alibaba had not disclosed parameter counts.

The benchmark claims come from Alibaba and have not been independently reproduced. Qwen4-Max scored 52.69 on Work360, a test Alibaba built covering 721 professional specialties, against 47.74 for Qwen3.8-Max. The company also reported gains on DeepSWE and HealthBench Professional. Its own comparison does not put the model first everywhere: on Terminal-Bench 4.0 it scored 48.5, behind Moonshot's Kimi K3.1 at 67.2 and Anthropic's Claude Opus 5.5 at 59.6.

Alibaba has also leaned on safety in its messaging. Training included security checks alongside coding tests, aimed at reducing flaws such as code injection and memory-safety bugs, and the company said it had tightened safeguards against misuse in areas including cybercrime and hazardous materials, Reuters reported. Alibaba said internal testing found no significant high-risk capabilities; no outside evaluation has been published.

The release gives enterprise buyers outside China another lower-priced option for agent and coding workloads, with the API available from data centre regions including Singapore, Hong Kong, Tokyo, Frankfurt and US Virginia. The model does not support fine-tuning or batch inference at launch. The Token Plan preview of the model runs until November 25.

By

Nakoda News AI & Tech Desk

AI models, chips, products and research, and what they can and cannot do.

About the desk

Related coverage