Alibaba Lists Qwen4-Max 25% Cheaper Than Its Predecessor in Singapore
The Takeaway: Alibaba’s new flagship model, Qwen4-Max, is listed 25% below its predecessor for standard input and output on QwenCloud’s Singapore price list, and its cached-input rates fall further. Its backbone size and context window match its predecessor’s listing, and Alibaba credits its gains to a redesign: on its own workplace test, the average cost per task fell 62%. Those efficiency figures come from tests the company ran itself, and Alibaba reports its September quarter results on Thursday.
Alibaba’s Qwen team released Qwen4-Max on 16 November, pairing a reworked architecture with lower prices on QwenCloud, where standard input costs $1.5 per million tokens on the Singapore list.
The model is built on the new Qwen4 architecture and adds native vision, the team said in a blog post. It is available on QwenCloud through the Token Plan or the API, the paid route developers use to call the model, and on QwenWork, Qoder and QwenStudio. The team, part of Alibaba Group (HKEX: 9988), calls it the first look at the Qwen4 family.
Prices fall on every rate the two model pages share
QwenCloud quotes its prices per million tokens, the units of text a model reads and writes. On its list for the Singapore region, standard input and output on the new model both cost 25% less than on its predecessor at the top of the family.
The cut is narrower on the US East list, where both models cost less: standard input and output on the new model are 14.4% cheaper than on its predecessor.
On the Singapore list, the deepest cuts are in cached input, text the system has already processed and keeps for reuse. The implicit cache rate falls 70% and the explicit cache read rate 55.9%. Both now stand at $0.075 per million tokens.
On the new Singapore list, a cached input token costs 5% of a fresh one, against 12.5% on the old list; the US East lists give the same proportions. Developers who send the same long documents or instructions again and again therefore save more than the headline cut.
QwenCloud list prices, previous and new flagship (in $ per million tokens)
| Rate | Previous | New | Change |
|---|---|---|---|
| Input | 2 | 1.5 | -25% |
| Output | 6 | 4.5 | -25% |
| Input, implicit cache | 0.25 | 0.075 | -70% |
| Explicit cache read | 0.17 | 0.075 | -55.9% |
| Five-minute cache creation | 2.5 | 1.875 | -25% |
Previous is Qwen3.8-Max and new is Qwen4-Max, as listed on the two model pages for the Singapore region on 17 November 2026. Changes are The Edge’s calculations.
The new model’s Singapore page also lists an explicit cache creation rate of $3 per million tokens for one hour, a rate the older page does not carry.
Capacity rises on the Singapore list as well: the new model allows up to 5 million tokens a minute, 2.5 times the older limit. On the US East pages, both models allow the same 5 million.
A same-size backbone works with fewer active parameters
The new model has 2.4 trillion backbone parameters, of which 71 billion are activated for each token. A new N-gram embedding layer, which looks up representations of local token patterns, adds 328 billion parameters.
The older model’s page describes it as a 2.4-trillion-parameter flagship as well, and both models carry a one-million-token context window, the amount of text a model can take in at once. Neither headline number is new to the family.
What changed is the design. The team says it upgraded its sparse mixture-of-experts model, a design that activates only part of its parameters for each token, across attention, residual connections and embeddings. Training at the full scale ran with no loss spikes, which the team attributes mainly to those changes and to training optimization.
On the team’s in-house tests of knowledge, code, question answering and science, technology, engineering and mathematics, the base model posts about 5% to 13% lower bits per byte than its predecessor’s base model, a measure of how well a model predicts text where lower is better. It does so while activating 25% fewer parameters.
The finished model also uses fewer tokens to complete the same work, the team said. On Terminal-Bench 4.0, a test of coding through a computer terminal, token use drops by about 48% at the Xhigh effort setting. At the High setting the cut is about 49%.
The team’s own workplace test, Work360, shows smaller savings, about 24% at Xhigh. At High the saving is about 35%.
Alibaba measures the gains on its own workplace test
The benchmark covers professional work across 26 top-level domains and their subfields, and the team built it itself. The new model scored 52.69 on it, against 47.74 for its predecessor.
On the same test, the average cost per task fell to $1.67 from $4.43. The team puts the reduction at 62%. The post reports that saving next to both the token savings and the price cuts, without splitting it between them.
Alibaba Cloud’s press release of the same day puts the cut in average cost per task on that test at around 60%, with 35% fewer tokens at the High setting. Both are test results, not changes to the list prices.
Alibaba also points to work outside the lab. In automotive production settings at SANF and Pacific Union, the manual effort to program a single workpiece fell to 24 minutes from about 475 in one showcase, the team said.
Inside Alibaba Cloud, the model has been put to work scheduling computing resources. The team’s post puts the fall in call scheduling costs for data annotation at an average of 10.42% against the reference baseline, and the gain in solving speed in offline CPU rescheduling at more than 14.61 times. Online GPU allocation increased whole-machine idle capacity duration by 4.62%.
On secure coding, the function-level pass rate for both functionality and security on Last Secure Code Bench rose to 31.8% from 22.0% for the previous generation.
The post also describes a 48-hour autonomous run in which Qwen4 designed a coolant molecule for data-center chips that an automated laboratory then synthesized and tested.
The rival comparisons and the revenue stay off the page
The post gives its figures for rival models only in chart images, and The Edge does not carry them. It says some of the scores in those charts are taken from outside leaderboards, among them Artificial Analysis and the Vals Index.
Alibaba gave no revenue, usage or customer figures for the model, in the blog post or in the press release. Between the evening of 16 November and the afternoon of 17 November, Hong Kong time, its only filing on the Hong Kong exchange covered a quarterly report on its holdings of US-listed equities.
As The Edge set out on 31 October, Apple said that day that it would add Alibaba’s Qwen model to Siri, Writing Tools and Visual Intelligence for iPhone users in China with iOS 27.3 in December.
In The Edge’s reading, the price list is the firmer news for developers: the cuts are posted on Alibaba’s own cloud pages and apply to every request billed at those rates, while the efficiency gains rest on tests the company ran itself. The team says smaller Qwen4 models, trained with the same techniques, are coming to open source; it gave no date.
What to watch:
Thursday 19 November: Alibaba Group reports its September quarter results before the US market opens, with a conference call at 12:00 GMT.
Model details, efficiency and cost figures are from the Qwen team’s blog post “Qwen4-Max: Your New Teammate Has Arrived” of 16 November 2026 and Alibaba Cloud’s press release of the same date; prices and limits are from the QwenCloud pages for the new and previous models, Singapore and US East regions, as read on 17 November 2026; results timing is from Alibaba Group’s announcement of 6 November 2026; percentages and ratios not attributed to Alibaba are The Edge’s calculations.
Sources: Alibaba Group, Hong Kong Exchanges and Clearing, Apple, The Edge.



