In brief. Alibaba reports a 62% drop in average cost per task, from $4.43 to $1.67, when moving to Qwen4-Max in its own Work360 test (report of 16 November 2026). The strategic point is to reassess the services and projects a company has ruled out because they cost too much.
$4.43. Then $1.67. That is the average cost per task Alibaba reports when moving from Qwen3.8-Max to the new Qwen4-Max in its own Work360 test: a 62% reduction. The Qwen Team report, published on 16 November 2026, attributes the result to fewer tokens used and lower prices. Tokens are the units of content the model processes and that add up to the bill.
It is the vendor’s comparison, on its own tests. No company can move that saving into its budget without verifying it. But it is the number to watch alongside the rankings of model intelligence. If the cost of getting a task done changes like this, it becomes necessary to reopen the drawer of projects that were discarded because they cost too much.
Part of the competition moves there: to the work that yesterday was not worth the cost.
The video precedent
My first startup, in 2001, was born from a conviction: broadband was coming, and it made sense to get ready to build projects with video. I have told this story when talking about what I consider a possible “YouTube moment” for software. When tools and distribution become accessible, who can afford to produce changes too.
The most interesting shift happens when cost stops excluding a whole category of ideas. A film can exist for a small audience. An application can serve a problem too specific to have justified an IT project before. Lowering the economic threshold changes demand, as well as supply.
With AI agents, that threshold concerns a growing amount of intellectual work. Preparing a comparison, checking documents, exploring a design variant: tasks that can be useful and still stay out of the plan, because someone has to spend time on them. Reducing their cost makes it possible to discuss again how many to do, for which clients and how often.
This is the limit of judging AI only by hours saved. That calculation sees the tasks that already exist very well. It struggles to see the ones nobody ever authorised.
A model enters the quote
Qwen4-Max is offered through a service that applications can use. The Alibaba Cloud documentation updated on 16 November describes text, image and video input, tool use and a window of up to one million tokens. These features make it possible to design work that runs across different documents and programs; they do not guarantee that the result is correct.
For a services company, the commercial question becomes concrete. How many alternatives can I prepare before delivering a proposal? Can I update an analysis when an assumption changes, instead of waiting for the next engagement? Can I serve a small request too, without it absorbing the whole margin?
These are economic possibilities to verify on your own work. The strategic news is that the boundary between a sustainable service and one that costs too much can move with every generation of the model. A task ruled out six months ago deserves a new assessment, even if the client’s need has stayed the same.
This changes the role of those who know the trade. Expertise also serves to recognise a demand left unanswered and to build a service around it. Knowing how to ask the machine for a document is only one step. Understanding who is willing to pay to have it updated, verified and usable requires a knowledge of the client that the model’s price list does not contain.
The cost left out of the chart
The strongest objection is economic: running the model can be the least expensive part. Retrieving usable data, connecting the tools and correcting errors can absorb the whole advantage. If a result that costs a few dollars requires a long professional review, the price of the service has not collapsed.
That is true. The cost to measure runs all the way to the accepted delivery. It includes failed attempts, corrections and checking time. For some jobs the new model will widen the market; for others it will change little. And an average improvement can coexist with rare errors costly enough to make automation not worth it.
The duration of the promise made to the client matters too. Selling an annual service on the basis of a model price that can be changed tomorrow exposes the margin. The commercial contract has to absorb that variability, or the company must be able to change supplier without rebuilding the whole service.
So the test has to be built around a precise result. The same request goes through the current process and the proposed one; you compare acceptable quality, full cost and elapsed time. Whoever judges must be able to reject a delivery, otherwise the reduction in spending simply hides incomplete work.
But stopping at the trial of the task that already exists leaves half the problem open. The comparison also needs a request we turn down today: a frequent update, a customisation for a few clients, a deeper analysis that always ends up below the priority line. It is on that request that we find out whether we have gained only margin or a new market.
Alibaba has not shown that every company can work while spending 62% less. It has published a signal strong enough to make some quotes old. Those who lead an organisation need to know where they are: which decisions not to do something still depend on a cost that has changed in the meantime.
Read on the site The test comes before the code · One department’s productivity, another’s queue
Impact deep dive Alibaba makes Qwen4-Max available pay-as-you-go (in Italian)