Industry News

Is AI Cost Per Task the Real Story Behind Qwen4-Max's Trending 60% Cut?

Alibaba says Qwen4-Max cuts average cost per task by about 60% against its predecessor. Independent testing confirms the economy and denies the supremacy: 43 on Artificial Analysis Intelligence Index, $3.27 per task, 32 minutes per task, 17 points behind Gemini 4 Carbon.

Alibaba released Qwen4-Max on November 16, and the headline number the company wants you to remember is not a parameter count or a benchmark rank. It is 60 percent: roughly how much the average cost of finishing a task fell compared with its predecessor, Qwen3.8-Max, on Alibaba's own professional-work benchmark.

That framing is genuinely new, and it is also where the argument starts. Because the moment you check the claim against an independent measurement, the 60 percent survives in one sense and collapses in another. Qwen4-Max really is cheaper per task than the model it replaces. It is also, by the only public apples-to-apples yardstick available, a mid-table model that consumes a lot of tokens and finishes tasks slowly. Both things are true at once, and understanding why is more useful than picking a side.

What Alibaba Actually Shipped

The specification sheet is straightforward, and the interesting detail is in how few parameters do the work.

  • Qwen4-Max carries 2.4 trillion total parameters but activates only 71 billion per token, 25 percent fewer activated parameters than Qwen3.8-Max at the same 2.4 trillion scale.
  • It handles native vision and a one-million-token context window.
  • It ranked sixth on Code Arena, a live coding leaderboard.
  • Input tokens that hit the prompt cache cost $0.075 per million, with output at $4.50 per million and standard input at $1.50 per million.
  • The model is proprietary. Alibaba says smaller open-weight models from the Qwen4 family will follow "soon," without dates.

The architectural story is a sparse mixture-of-experts rebuilt across attention, residual connections, and embeddings, plus a new sparse attention scheme aimed at long-context cost. Alibaba describes the redesign as the source of both the capability gain and the cheaper inference; it does not publish a single clean multiplier for long-sequence speed, so treat any "N times faster" figure you see quoted elsewhere as a secondary claim rather than a released number.

What the release does quantify precisely is the cost argument, and it does so in three places.

The 60 Percent Claim, and What It Is Measured Against

Alibaba's number comes from Work360, an internal benchmark covering 721 specialized professional fields. On that benchmark, Qwen4-Max beats Qwen3.8-Max overall while using 35 percent fewer tokens at the High reasoning setting and cutting the average cost per task by around 60 percent.

Read that sentence carefully. It compares Qwen4-Max to Qwen3.8-Max. It does not compare Qwen4-Max to Claude, Gemini, or Kimi, and it does not use a public benchmark anyone else can rerun. The improvement is real in the sense that Alibaba measured it and published the method; it is not evidence of a market-wide cost lead.

Two supporting figures from the same release point the same direction. In cloud resource scheduling, where the model optimizes its own workflows in a closed loop, Alibaba reports annotation model call costs down an average of 9 percent and offline CPU rescheduling more than 14 times faster. In automotive manufacturing, a showcase cut manual workpiece programming from roughly 475 minutes to 24 minutes, a 95 percent reduction, by catching overcuts and insufficient material removal before machining started.

Efficiency claims of this shape are usually about token discipline rather than raw intelligence, and that distinction matters for anyone budgeting an agent.

Independent Testing Puts the Model in the Middle of the Pack

Artificial Analysis, which runs a fixed nine-evaluation suite across models and publishes cost and latency alongside quality, scored Qwen4-Max at 43 on its Intelligence Index. That places it 38th of 81 models measured, just above the median of 41 for comparable models and 16 points behind Claude Opus 5.5 at 59.

The cost side of the same page is more flattering. Artificial Analysis measures Qwen4-Max at $3.27 per Intelligence Index task, which ranks 44th of 80 on cost — good, but not exceptional. The measurements are as follows:

  • Qwen4-Max: index 43, $3.27 per task, 32.2 minutes per task, 54 output tokens per second.
  • Kimi K3.1 Pro: index 46, $3.68 per task, 38.5 minutes per task, 31 tokens per second.
  • Gemini 4 Carbon: index 60, $5.48 per task, 31.9 minutes per task.
  • Claude Opus 5.5: index 59, $10.04 per task, 24.6 minutes per task.

Three observations fall out of that list. First, Qwen4-Max is about 67 percent cheaper per task than Opus 5.5 and roughly 40 percent cheaper than Gemini 4 Carbon, while giving up 16 and 17 index points respectively. Second, against Kimi K3.1 Pro it is only about 11 percent cheaper per task and three points less intelligent, which is a much narrower advantage than the launch framing suggests. Third, it is not fast: 32.2 minutes per task against a 13.3-minute median, with 179.8 seconds to the first answer token.

Divide quality by measured cost and the picture sharpens. Qwen4-Max returns about 13.1 index points per dollar of task cost, Kimi about 12.5, Gemini 4 Carbon about 10.9, and Opus 5.5 about 5.9. On that ratio, Alibaba's model is the best value in this comparison and also the least capable of the four. That is not a contradiction. It is what a budget frontier model looks like.

The Verbosity Tax Nobody Advertises

The most underrated number on the Artificial Analysis page is token consumption. To complete the Intelligence Index, Qwen4-Max generated 127 million output tokens. The median model in its comparison class generated 78.6 million. It also read 9.9 billion input tokens across the suite.

This is the mechanism behind the cost story, and it cuts both ways. A model that thinks longer can post better scores while spending more; a model that is tuned to think less can cut cost per task while leaving accuracy on the table. Alibaba's own Work360 claim of 35 percent fewer tokens at the same effort setting suggests the company is deliberately trading reasoning length for economy. The independent data suggests it succeeded at the economy and landed in the middle on quality.

For a team running thousands of agent calls a day, that trade is entirely rational. For a team running a handful of hard research tasks a week, three points of index and 18 extra minutes per task may matter more than a few dollars.

Where the Efficiency Shows Up in Practice

Alibaba attached two long-horizon demonstrations to the release, and they are more informative than the benchmark table because they show what "cost per task" means when the task is measured in hours.

In a chip-cooling problem, Qwen4 ran a continuous 48-hour agent loop through four iterations of a closed workflow — molecular design, synthesis and testing, active learning — and arrived at a coolant molecule that was experimentally validated. In an Alzheimer's research workflow, the model handled target research, molecular design, and computational evaluation end to end, executing 454 tool calls and 3,239 metric evaluations to design and optimize 24 candidate molecules.

Neither result is a claim about scientific novelty; both are claims about sustained execution without a human babysitting each step. That is the actual product being sold here: not a smarter model, a cheaper one that can be left running.

The release also notes a safety posture worth recording. Alibaba says software-engineering training included code security checks alongside functional tests, that training data was filtered by task intent to keep defensive vulnerability material and drop offensive attack content, and that dedicated datasets and post-training target chemical, biological, radiological, and nuclear risk. These are self-reported process claims, not audited outcomes.

What This Means If You Are Actually Buying

The practical lesson from Qwen4-Max is not "switch to Qwen4-Max." It is that the unit you compare has changed, and most procurement conversations have not caught up.

Price per million tokens is a poor proxy for what a model costs you, because it says nothing about how many tokens the model spends getting to an answer. Qwen4-Max at $1.50 input looks cheaper than Kimi at $2.00, and it is, but the gap in measured cost per completed task is 11 percent, not 25 percent, because the models consume different amounts of reasoning. Multiply that by an agent that loops 40 times per ticket and the difference between a good per-token price and a good per-task price becomes the difference between a working budget and a surprise invoice.

This is the reasoning behind how AI Bookmarks organizes its directory: tools and models are grouped by what they actually do and what they cost to run a job, not by who topped this month's leaderboard. A model that ranks 38th and finishes a task for three dollars is the right answer for a batch pipeline and the wrong answer for a one-shot legal review, and no single ranking number conveys that. The site's comparison pages put measured cost, latency, and context limits side by side for exactly that reason.

The Open-Weight Question Hangs Over All of It

Qwen4-Max is closed. No weights, no fine-tuning, no batch inference, no dedicated deployment on Model Studio as of launch. Alibaba's promise of open-weight smaller Qwen4 models is the part of the announcement with the longest tail, because the efficiency tricks in the flagship — sparse attention, the 71-billion activation budget, the cache economics — are the tricks that matter most at small scale.

If those land, the cost-per-task argument stops being a choice between closed frontier models and becomes a choice about where you run the weights. That is the scenario worth watching over the next quarter, and it is a better use of attention than another leaderboard screenshot.

Bottom Line

Qwen4-Max is a competent, deliberately economical model that Alibaba has positioned with an unusually honest metric. The 60 percent cost reduction is measured against its own predecessor on its own benchmark, and it holds up as a statement about token discipline. Independent testing confirms the economy and denies the supremacy: 43 on the Intelligence Index, $3.27 per task, 32 minutes per task, 17 points behind Gemini 4 Carbon.

  • If your workload is high-volume, tolerant of mid-tier accuracy, and billable per task, Qwen4-Max is competitive with anything in its price band, and slightly ahead of Kimi on measured cost.
  • If your workload is low-volume and high-stakes, the 16-point gap to Opus 5.5 will cost you more in review time than you save in tokens.
  • If you are still comparing models by price per million tokens, you are optimizing the wrong variable.

The most useful thing Alibaba did this week was not ship a 2.4-trillion-parameter model. It was to make cost per task the number everyone now has to argue about.

---

Compiled by the AI Bookmarks editorial desk from Alibaba Cloud's November 16, 2026 release and Artificial Analysis Intelligence Index v5.0.1 measurements. Vendor-reported figures are identified as such; index scores, per-task costs, and latency figures are independent measurements.

Published by AI Bookmarks · More AI news →

← Back to News