Alibaba Qwen4-Max Is Generally Available: 2.4T MoE, 1M Context, $1.50 per Million Tokens

Alibaba's Qwen4-Max is generally available from today, 16 November 2026, as a pay-as-you-go API with model ID qwen4-max, priced at $1.50 per million input tokens and $4.50 per million output tokens in Singapore. It is a sparse mixture-of-experts flagship with a 2.4-trillion-parameter backbone, only about 71 billion parameters active per token, and a 1-million-token context window that accepts text, image and video input.

Its launch arrived the same day through Reuters, which reported that Alibaba says the model outperforms its predecessor, Qwen3.8-Max, while using fewer tokens and costing less per task.

What Qwen4-Max actually is

A mixture-of-experts model, or MoE, is a model that keeps a huge total parameter count but activates only a small subset of those parameters for each token it processes, so it can be both very large and relatively cheap to run. Qwen4-Max follows that design: a 2.4-trillion-parameter backbone with about 71 billion parameters activated per token, plus a separate 328 billion parameters of N-gram Embedding. Alibaba says the new architecture activates 25% fewer parameters than the Qwen3.8-Max base model while scoring better in every domain it measured.

The headline specifications, all from Alibaba's own QwenCloud model page for the Singapore region, are straightforward. The context window is 1 million tokens, with a maximum input of 991,000 tokens and a maximum output of 393,216 tokens. Input is text, image and video; output is text. The model supports function calling, structured outputs, context caching and web search, and it works through OpenAI-compatible and Anthropic-compatible APIs as well as Alibaba's own DashScope interface.

Developer working at a desk with code on a monitor and an AI chat panel open, photorealistic
AI-generated illustration (generated 16 Nov 2026).


How much Qwen4-Max costs

The pay-as-you-go price is $1.50 per million input tokens and $4.50 per million output tokens. Cached input is far cheaper at $0.075 per million tokens, which matters more than the headline rate for agent work, because agents resend the same context on every turn. Explicit cache creation costs $1.875 (five minutes) or $3.00 (one hour) per million tokens.

ModelInput per 1M tokensOutput per 1M tokensCached input per 1M tokens
Qwen4-Max$1.50$4.50$0.075
Qwen3.8-Max$2.00$6.00$0.25
Claude Opus 5.5$4.00$20.00—

The gap is widest on output tokens, where Qwen4-Max costs less than a quarter of Claude Opus 5.5's $20 per million. Price is not quality, and a cheaper model that gives worse answers is no bargain at all, which is why the benchmark question matters. But on price alone, Qwen4-Max is now the cheapest flagship-class API in this flagship AI model comparison. For context, Google's Gemini 4 Carbon and OpenAI's GPT-6.1 Terra launched at very different price points, and the Claude Haiku 5.5 price cut showed Anthropic is also competing hard on cost at the lighter end of its range.

Preview versus general availability: do not mix them up

The release came in two steps, and confusing them is the easiest mistake a developer can make this week. On 11 November 2026 the official Qwen account announced Qwen4-Max-Preview, model ID qwen4-max-preview, reachable only through Alibaba's Token Plan subscription and Alibaba's own apps. The preview window runs from 10:00 on 11 November to 24:00 on 25 November 2026 (UTC+8), and Alibaba's own documentation says the preview model will be retired after that window ends.

General availability is a different model ID on a different billing route. The generally available model is qwen4-max, served on the QwenCloud pay-per-token API. The Token Plan page on 16 November still lists both IDs, with the preview tagged as a limited-time offer, but new production work belongs on qwen4-max. Anyone who wired qwen4-max-preview into a coding agent during the preview should switch the model ID before 25 November, because that ID is scheduled to disappear. The Token Plan itself is a monthly subscription that buys a pool of Credits for interactive use inside coding tools such as Claude Code, Cursor and Qwen Code; its own help pages say it is for interactive use in compatible programming and agent tools only and cannot be used for automated scripts or application backends. For backends, use the pay-as-you-go API key, not a Token Plan key.

Modern data centre corridor with rows of server racks under cool blue light, photorealistic
AI-generated illustration (generated 16 Nov 2026).


What Alibaba claims it can do — and what is still unproven

Alibaba did publish benchmarks with the launch, and they need to be read as the vendor's own figures. On Work360, Alibaba's own benchmark of industry tasks across 26 domains, Qwen4-Max scores 52.69, up from 47.74 for Qwen3.8-Max. Token use falls about 24% at extra-high effort and 35% at high effort, and about 48% to 49% on Terminal-Bench 4.0. Combined with the lower prices, the average cost per Work360 task falls from $4.43 to $1.67, a 62% cut. The launch post also includes a 16-benchmark chart comparing Qwen4-Max with Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Moonshot's Kimi K3.1.

Those numbers are Alibaba's own measurements, and no independent lab had published a full rerun on launch day. Reuters noted the same gap, reporting that the announcement lacked comprehensive data to independently verify the model's claimed capabilities. There is, however, one same-day independent data point: Artificial Analysis scored Qwen4-Max 43 on its Intelligence Index, above its predecessor, and reported a cost of $3.27 per index task against $10.02 for Qwen3.8-Max. Speed and latency data were not available on launch day, so response time cannot yet be compared.

Alibaba also puts unusual emphasis on safety in this launch. The company says training included security checks alongside coding tests. Alibaba also says its in-house security experts found no evidence that the model could pose large-scale cybersecurity risks. Those are the company's own claims, reported by Reuters, and they have not been independently verified either.

Where you can use it, and who should try it first

Alibaba lists five routes to the model. The QwenCloud API is the production route for app backends and scripts. The Token Plan is for driving a coding agent by hand. QwenWork is Alibaba's AI productivity app for turning a goal into reports, slides or web pages. Qoder is its agentic coding platform, and QwenStudio is the chat interface for trying prompts without setup. The launch post also shows the model working in Blender, Unity, Unreal Engine and Godot, and driving a 48-hour autonomous laboratory run that designed and synthesised a new coolant molecule. Those are Alibaba's own showcases, not independent case studies.

The most important limitation is openness. Qwen4-Max itself is not open weight: there are no downloadable weights, and it runs only on Alibaba's servers. That breaks with the Qwen tradition, because the Qwen3 family is the most downloaded open-weight model family in the world. Alibaba says the open weights of smaller Qwen4 models will be released, with no names and no dates. If you need to run a model on your own hardware, Qwen4-Max is not that model. If you want a cheap flagship-class API with a million-token context, it is worth a test: start on the pay-as-you-go API with model ID qwen4-max, test it on a real task you can grade objectively, and compare the cost and the result with the model you use today. Vendor benchmarks point in a direction; your own tasks settle it.

Close-up of a laptop screen showing a long document being summarised by an AI assistant, photorealistic
AI-generated illustration (generated 16 Nov 2026).


Key Takeaways

  • Qwen4-Max is generally available from 16 November 2026 as qwen4-max on QwenCloud pay-as-you-go, at $1.50 input and $4.50 output per million tokens, with cached input at $0.075.
  • It is a sparse MoE flagship: a 2.4-trillion-parameter backbone, about 71 billion active per token, a 1-million-token context, and text, image and video input.
  • The preview ID qwen4-max-preview was Token Plan only from 11 November and retires on 25 November 2026 — switch production work to qwen4-max before then.
  • Alibaba's own figures claim a 62% cut in average cost per Work360 task, from $4.43 to $1.67; these are vendor figures, not independent tests.
  • Artificial Analysis independently scored Qwen4-Max 43 on launch day, above its predecessor, but speed data was unavailable.
  • There are no open weights for Qwen4-Max; smaller Qwen4 models are promised open source, with no date given.

Conclusion

Qwen4-Max is a serious launch, not a rebrand. A flagship with a million-token context at $1.50 and $4.50 per million tokens changes the price floor for top-tier models, and the 62% cost-per-task claim, if it holds up on real work, is the number that will move developers. But every performance figure so far comes from Alibaba itself, and the preview ID disappears on 25 November. Try it on a task you can grade, and switch the model ID before the preview retires.

Sources

 

Comments