A cheaper flagship built for longer tasks
Alibaba has made Qwen4-Max generally available through QwenCloud, pairing a context window of one million tokens with lower prices and a claimed reduction of about 58% in the average cost of completing tasks on its Work360 benchmark. The launch follows a limited preview that opened on November 11, 2026.
Contents
- A cheaper flagship built for longer tasks
- From subscription preview to wider access
- What fits inside a million-token window?
- How the price reduction works
- Repeated documents can cost less to process
- A large model that activates a smaller share
- Benchmarks show gains and uneven competition
- What developers can connect today
- Agents gain a revision step
- Key Points
The flagship model uses a mixture-of-experts architecture with a backbone of 2.4 trillion parameters. It accepts text, images and video, then produces text. Alibaba is positioning it for coding, office work, frontend design and agents, meaning software that can use tools and carry out a sequence of steps toward a goal.
Alibaba's Model Studio documentation, updated on November 16, lists international prices of $1.50 per million input tokens and $4.50 per million output tokens. Tokens are the small pieces of information a model processes. For text, a token can be a word, part of a word or punctuation.
Alibaba says Qwen4-Max also needs fewer tokens than Qwen3.8-Max for comparable work. Lower prices and reduced usage together brought the average Work360 task cost from RMB 26.58 to RMB 11.16, a saving of RMB 15.42. Another published dollar calculation gives $4.43 to $1.67, or about 62%; the renminbi figures yield the roughly 58% reduction cited in the launch account.
For developers, the practical attraction is the combination: more room for documents and conversation history, access to visual material, and a lower charge for each unit of processing. The size of the saving on an actual application will depend on its requests, reasoning settings and generated answers.
From subscription preview to wider access
Qwen4-Max first appeared through Alibaba's Token Plan subscription. Early access also reached the Qwen app, Qoder, QwenWork and QwenStudio. General availability expands access through QwenCloud, while the model remains available in Alibaba's work and development applications.
The published sequence is short:
- November 11, 2026: Alibaba opened the limited preview, with developer access through Token Plan and early use through its applications.
- November 16, 2026: Alibaba updated the Model Studio documentation with capabilities, prices, context limits and API details.
- General availability: Alibaba opened QwenCloud access after the preview; an exact launch date was not given.
Developers call the model using the identifier qwen4-max. Model Studio also lists a free quota of one million tokens, valid for 90 days from activation, model release or application approval, whichever comes later.
What fits inside a million-token window?
Qwen4-Max's context window is the space available for information involved in a request. A larger window lets an application bring more documents, code or conversation history into the same task, rather than dividing everything into smaller requests.
Alibaba specifies maximum input lengths of 991,808 tokens in non-thinking mode and 983,616 tokens in thinking mode. The two modes offer different ways of handling a task: thinking mode gives the model a reasoning budget before it delivers its answer.
The documentation lists a maximum output of 131,072 tokens in non-thinking mode and a combined limit of 262,144 tokens for thinking and the final answer. A published launch account gives a higher maximum output figure of about 393,000 tokens; Alibaba's current service documentation provides the operating limits developers can configure against.
Text, image and video inputs broaden the kinds of material an application can submit. A work request could, for example, involve written instructions alongside screenshots or video. The documented output format is text, including code and structured data.
How the price reduction works
International API prices fall from the reported Qwen3.8-Max rates of $2 per million input tokens and $6 per million output tokens to $1.50 and $4.50 respectively. Both changes are reductions of 25%. Alibaba's documentation separately lists global deployment prices of $1.413 for input and $4.24 for output per million tokens.
A request using 100,000 input tokens and producing 10,000 output tokens would cost $0.195 at the international rates, before any cache savings or other applicable charges. The same quantities at the earlier $2 and $6 rates would cost $0.26. This calculation holds usage steady to show the effect of the price change alone.
Alibaba's larger claimed task savings also depend on using fewer tokens. On Work360, reported token consumption fell by about 24% at the Xhigh reasoning setting and 35% at High. On Terminal-Bench 4.0, the reported reductions were about 48% and 49% respectively.
Reasoning is part of the bill: the pricing table describes output charges as covering both chain of thought and the answer. An application that asks for extensive reasoning can therefore spend more than one requesting a short response, even when the user sees a similarly brief final answer.
Repeated documents can cost less to process
Qwen4-Max supports explicit and implicit caching. A cache lets a service reuse previously processed input, such as a standard instruction, a tool description or a reference document, rather than charging the ordinary input rate for the same material each time.
Alibaba's documentation says explicit caching works with OpenAI Chat Completions, OpenAI Responses and Anthropic Messages interfaces. It lists cache lifetimes of five minutes and one hour. The reported implicit cache read price is $0.075 per million tokens, down from $0.25, a reduction of 70%.
Repeated use matters for agents because they may consult the same instructions and documents across several steps. A lower token price, fewer generated tokens and reuse of common input can all contribute to the cost of finishing a job. Their contribution will vary with how often an application repeats information and how much work the model performs.
A large model that activates a smaller share
Qwen4-Max's reported 2.4 trillion backbone parameters describe the numerical settings learned during training. Its mixture-of-experts design divides some of that capacity among specialist components and selects a subset as it processes each token.
The architecture details put active parameters at 71 billion per token, roughly 3% of the 2.4 trillion backbone. They also describe another 328 billion parameters used for N-gram embeddings, which store representations of short, local token patterns.
Alibaba's design combines Qwen Sparse Attention with Gated DeltaNet layers. In simple terms, some layers compress earlier information into a compact state, while an indexing mechanism selects relevant portions of the longer context. This is intended to reduce the processing required for long sequences.
Gated Residuals provide four parallel routes for information, with gates controlling reading and writing. Alibaba reports stable training without loss spikes and says its base model achieved 5% to 13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters. Bits per byte measures prediction efficiency during evaluation; a lower value means the model predicts the test text more efficiently.
Benchmarks show gains and uneven competition
Alibaba's comparison chart places Qwen4-Max first on four of 16 benchmarks. Its strongest listed advantage is on HLAB-Vals-1.1, a legal agent evaluation, where it scores 21.3 against 11.7 for Kimi K3.1 and 9.2 for Claude Opus 5.5.
The model also leads the chart on Agents' Last Exam v1.1 at 46.7, BioMysteryBench at 85.0 and DeepSWE v1.1 at 75.7. The margins in the latter two are narrow: GPT-6.1 Sol scores 84.4 on BioMysteryBench, while Kimi K3.1 scores 75.5 on DeepSWE.
Alibaba's Work360 result rises from 47.74 for Qwen3.8-Max to 52.69 for Qwen4-Max, an increase of 4.95 points. Work360 spans 26 top level domains and 721 specialized fields, linking the company's performance and cost claims to a broad collection of work tasks.
Other results show larger gaps behind competitors. Qwen4-Max scores 31.5 on ProgramBench against 65.0 for Claude Opus 5.5, and 54.2 on Chartography against 71.1 for GPT-6 Astra. Claude Opus 5.5 leads seven of the 16 charted evaluations. The launch comparisons combine Alibaba's results with some scores drawn from external leaderboards, without independent reproduction at launch.
What developers can connect today
Alibaba supports OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope APIs. Compatibility with these interfaces gives developers familiar ways to send requests and receive results, although applications still need configuration and testing for the new model.
The documented features include streaming, web search, parallel tool calling and structured output using JSON objects or JSON Schema. Streaming delivers an answer as it is generated. Parallel tool calling lets the model request several operations at once, while structured output helps software receive information in a predictable format.
Thinking mode is enabled by default, with reasoning effort set to High and a thinking budget of 131,072 tokens. Supported effort levels are Low, High and Xhigh. Developers can adjust these controls to suit tasks with different demands for speed, cost and reasoning.
Alibaba's official capability table marks batch inference, fine-tuning and model deployment as unsupported. These service specifications are narrower than a published launch description listing batch processing and fine-tuning among supported features. Global limits include 30,000 calls and five million input and output tokens per minute, with either threshold able to restrict traffic.
Agents gain a revision step
Alibaba also released a plugin that lets agents deliver an initial version of their work and revise it in response to user feedback. This introduces a review stage into a workflow that might otherwise aim to finish a task in a single run.
A user can assess an initial result before asking the agent to change it. For office work, coding and design, that provides a concrete point for correcting direction, rather than waiting until a longer sequence of actions has finished.
Qwen4-Max itself is proprietary and available through hosted services, with no downloadable model weights for running it on a customer's own hardware. Alibaba plans to release smaller Qwen4 models as open source. For now, access to the flagship runs through the company's cloud and application products.
Key Points
- Qwen4-Max is generally available through QwenCloud after a preview that began on November 11, 2026.
- The model accepts text, images and video and supports a context window of one million tokens.
- International API prices are $1.50 per million input tokens and $4.50 per million output tokens, 25% below the reported predecessor rates.
- Alibaba says average Work360 task costs fell from RMB 26.58 to RMB 11.16, about 58%.
- Official documentation lists a maximum non-thinking output of 131,072 tokens and a combined thinking and answer limit of 262,144 tokens.
- Alibaba's chart puts the model first on four of 16 benchmarks, with weaker results on some coding and visual tasks.
- The flagship remains hosted and proprietary; smaller Qwen4 models are planned for open source release.






