Alibaba’s Qwen4-Max cuts AI task costs, but benchmarks show an uneven challenge to frontier rivals


Credit: VentureBeat made with OpenAI ChatGPT-Images-2.6
Alibaba says its newest leading AI model is meant to help businesses complete more work for less money. It may not rank first in overall intelligence, but it could still be useful for certain jobs.
In its Qwen4-Max announcement, Alibaba says the model can take in text, images, and video. It can work with outside software and handle up to one million tokens of context. Alibaba says Qwen4-Max is better than Qwen3.8-Max and lowers the average cost of a task in its own Work360 test from $4.43 to $1.67. That is a reduction of about 62%.
Qwen4-Max is a proprietary model. It is not an open-source or open-weights release. Artificial Analysis classifies it as proprietary, and Alibaba has not made its model weights available for businesses to download and run on their own computers. Customers can use it through the QwenCloud API, or through Qwen Studio, QwenWork, and Qoder.
This is an important change for a model family known for releasing models that developers can download and adapt. Alibaba's launch announcement promises open releases of smaller Qwen4 models, but gives no release dates or licenses. It makes no similar promise for Qwen4-Max. At the time of publication, we found no public promise to release Qwen4-Max's weights. Alibaba also has not said for certain that they will always remain closed. The flagship model is proprietary today; the announced open releases are for smaller models.
The API costs $1.50 for every million input tokens and $4.50 for every million output tokens. These prices are lower than those of several well-known models from Anthropic, OpenAI, and Google. However, token prices alone do not tell businesses how much it will cost to finish a job.
The performance results are more complicated. Qwen's launch chart shows competitive scores on some professional-work and long-task tests, but much lower scores on other coding and visual tasks. An independent overall test also places the model behind leading American competitors and Moonshot AI's Kimi K3.1 Pro.
For businesses, Qwen4-Max offers a possible choice for certain kinds of work. Its costs and features may be attractive, but buyers also need to consider its accuracy, speed, and limits on how it can be used.
Lower prices and fewer tokens make the model more efficient
Alibaba's Work360 framework covers professional work in 26 broad areas, 182 smaller areas, and 721 specialist fields. Examples include accounting reviews, manufacturing analysis, legal work, and document creation.
Qwen4-Max scores 52.69 on this test, compared with 47.74 for Qwen3.8-Max. That is an increase of 4.95 points, or about 10.4% compared with the older model.
Alibaba says Qwen4-Max uses about 24% fewer tokens at the Xhigh reasoning setting and 35% fewer at High on Work360. On Terminal-Bench 4.0, the reductions are close to 48% and 49%.
These results measure different things. The roughly 50% reduction in token use comes from the terminal test. The 62% reduction in cost comes from Work360 and includes both lower token use and lower prices. Neither result proves that every customer will save the same amount.
There is also independent evidence that Qwen4-Max can cost less to use. Artificial Analysis's Qwen4-Max evaluation reports a cost of $3.27 for each task in its Intelligence Index test. Its Kimi K3.1 Pro evaluation reports $3.68. This makes Qwen about 11% cheaper by that measure. These numbers come from a different test system than Alibaba's $1.67 result, so they should not be treated as directly comparable.
How API prices compare
The prices below are standard API prices in US dollars for every million tokens. Qwen's prices come from its Singapore-region listing, while Google's are global prices. Discounts for large batches, tool charges, regional price differences, and business agreements can change the final cost.
Model and pricing source | Input | Output | Cached input | Important qualification |
|---|---|---|---|---|
$1.50 | $4.50 | $0.075 | Cached rate applies to qualifying cache hits | |
$4.00 | $20.00 | $0.20 | Standard Claude API rates | |
$2.00 | $10.00 | $0.10 | Standard global rates | |
$2.00 | $10.00 | $0.10 | Above 272K input tokens, input/cache rates double and output rises to $15 | |
$2.00 | $10.00 | $0.10 | Introductory pricing through November 30; then $4/$20/$0.20 | |
$1.50 | $7.50 | $0.075 | Standard global rates | |
$2.00 | $10.00 | $0.20 | Rates reported by Artificial Analysis |
Qwen's output price is 55% lower than the $10 rate for standard Sol, Sonnet, and introductory Carbon use. It is 77.5% lower than Opus 5.5. Its input price matches Hydrogen's, while its output price is 40% lower.
However, Qwen4-Max is not the cheapest choice for every job. Anthropic's price list shows that Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts of up to 100,000 tokens. Above that limit, the prices rise to $0.50 and $2.50. Buyers still need to find out which model can complete their work reliably.
Qwen also charges separately for creating a cache. This costs $1.875 per million tokens for a five-minute cache and $3 for a one-hour cache. Reading from the cache costs $0.075. The low cached-input price does not apply to every token in a request.
Its separate Token Plan subscription advertises special starting prices from $3 for new customers. This offer, and a separate discount for qwen4-max-preview, should not be confused with standard Qwen4-Max API pricing.
Agent benchmarks show strengths—and sharp limits
An AI agent must do more than give an answer that sounds right. It needs to use tools, understand the results, change its approach when needed, and produce something people can use. Alibaba's launch comparison tests several parts of this process, and the results vary.
Benchmark | Qwen4-Max | Qwen3.8-Max | Opus 5.5 | GPT-6.1 Sol | Kimi K3.1 |
|---|---|---|---|---|---|
DeepSWE v1.1: coding agents | 75.7 | 69.3 | 74.2 | 75.2 | 75.5 |
Terminal-Bench 4.0: terminal work | 48.5 | 38.9 | 59.6 | 56.1 | 67.2 |
ProgramBench: software replication | 31.5 | 28.0 | 65.0 | 52.5 | 25.0 |
APEX-Agents v1.1: professional services | 72.1 | 63.3 | 73.5 | 60.0 | 65.5 |
AutomationBench-AA: software workflows | 72.5 | 56.2 | 69.5 | 64.9 | 73.3 |
ChartoGraphy: chart interpretation | 54.2 | 42.0 | 66.3 | 67.8 | 50.0 |
These numbers are taken from Qwen's chart, including the name it uses for Kimi. They have not been independently tested again. Alibaba's notes say that some scores came from outside evaluators, including Mercor and Artificial Analysis. Other results use specific settings or versions of test data.
Qwen narrowly gets the highest score among the models shown on DeepSWE and comes close to Opus on APEX-Agents. It also scores 46.7 on Agents' Last Exam v1.1, compared with 27.9 for its predecessor and 43.4 for Opus.
However, Qwen's Terminal-Bench score is far below Kimi's, and its ProgramBench score is less than half of Opus's. A business testing an AI system that handles software maintenance on its own should not rely on one good coding score. It needs evidence that the system can handle all the tasks it will be given.
The gap in chart reading also matters for work involving finance and document analysis.
Overall intelligence remains behind the leaders
Artificial Analysis's launch chart gives Qwen4-Max a score of 43 on its Intelligence Index, up from 34 for Qwen3.8-Max. Kimi K3.1 Pro scores 46. GPT-6.1 Sol and Claude Sonnet 5.5 score 55. Opus 5.5 scores 59, Gemini 4 Carbon scores 60, and Claude Fable 5.2 leads this group with 65.
These figures show the reasoning settings used in the chart. They do not necessarily show how each model performs with its default settings.
This overall score gives a fuller picture than Alibaba's selected task results alone. Qwen can perform almost as well as its competitors on certain jobs while still falling well behind them in the combined test.
Qwen4-Max is also somewhat slow. At the time of this review, Artificial Analysis's model page reports that it takes about 32 minutes to complete each Intelligence Index task and about 159 seconds to produce its first answer token. These times put it among the slower major AI models.
These measurements come from a demanding test and the particular model setup used for it. They do not tell us how long every normal response will take. Still, they show that lower prices do not always mean faster results.
Multimodal input, but not native media generation
The QwenCloud specifications say that Qwen4-Max can take in text, images, and video and produce text. They do not list audio input or direct creation of images, audio, or video.
This helps explain the demonstrations shown at launch. Qwen uses Blender for 3D work, game engines for interactive software, and WonderClip in a video-production process. Making a video with software tools is different from having the AI model create a video directly.
Other models have different abilities. Opus 5.5 accepts text and images and produces text. GPT-6.1 Sol also lists text and image input. Audio and video are not supported as native input types, although it can use tools such as image generation. Google's Carbon and Hydrogen models accept text, images, video, and audio and produce text. Artificial Analysis lists Kimi K3.1 Pro as accepting text and images.
For applications that mainly work with recorded speech, Qwen's published specifications mean businesses need a different way to connect audio to the model than they would with Google's audio-capable models. Qwen can directly accept images and video for visual checks and analysis, but businesses still need to test its accuracy and processing limits.
A larger architecture aims to reduce active computation
Alibaba describes a 2.4-trillion-parameter backbone. It activates 71 billion parameters for each token, with another 328 billion parameters used for local token-pattern representations. The 2.4 trillion figure does not include this additional component.
The model uses a design called a mixture of experts. This means it activates only some parts of its network at a time. It also uses compressed information from earlier text and selective attention to make long inputs cheaper to handle. Changes to the way information moves through the network are meant to make training more stable.
Alibaba says Qwen4-Max activates about 25% fewer parameters than Qwen3.8-Max. Its pretraining chart also shows better prediction scores for knowledge, code, question answering, and STEM. These scores measure how well the model predicts information, not how well it completes real tasks. Alibaba also notes that the comparison includes both basic models and models that received additional training.
The main practical question is whether these design changes help the model complete useful work reliably at an acceptable cost. The number of parameters alone cannot answer that question.
Reactions show different ideas about what makes a successful release
The launch has attracted criticism because some strong results on selected tests do not match the model's lower scores in broader rankings. In a post shown in a screenshot reviewed for this article, the X account Chubby, @kimmonismus, writes: “Qwen 4 is not a good release.”
The account argues that the Artificial Analysis score challenges claims that Qwen performs as well as Opus. It also suggests that restrictions on distillation explain the weaker results. However, the post provides no evidence for this explanation. Benchmark scores alone cannot show how access to other models during training affected Qwen's performance.
Independent cost measurements support a more balanced view. Qwen scores below Kimi on overall intelligence but costs less for each tested task. This could matter to a business processing many tasks, even if Qwen is not the leading model overall.
Availability, open weights, and the enterprise decision
Qwen4-Max is available through QwenCloud, Qwen Studio, QwenWork, and Qoder. Its API supports tool calling, structured responses, web search, batch processing, and formats that work with OpenAI and Anthropic APIs.
Alibaba also describes a Live Delivery workflow. It shows useful intermediate results while an AI agent is still working, so users can change its direction before it finishes. The company says it is releasing the related plugin as open source.
However, unlike many earlier Qwen models, Qwen4-Max does not have open weights. Alibaba promises downloadable weights for smaller Qwen4 models, but has not announced their licenses or release dates. Businesses that want to run the flagship model on their own computers cannot treat this future promise as something available today.
Alibaba's manufacturing and science demonstrations aim to show what the model can do. The company says it reduced the manual programming time for one machining workpiece from about 475 minutes to 24 minutes. It also describes a 48-hour materials-research project that produced a physical coolant sample. These are demonstrations reported by Alibaba, not proof that similar savings or discoveries will happen in other customer projects.
The clearest reason for businesses to consider Qwen4-Max is measurable: they should test whether its lower prices and improved task performance make it cheaper to produce acceptable results. The calculation should include repeated attempts, human review time, outside tools, and waiting time. Its strongest benchmark results make it worth testing, while its weaker results show why businesses must check how it performs on the exact work they plan to give it.
