MeanCEO: Tech Startups and Startup Ideas

Qwen4-Max: Alibaba's New Flagship Costs $1.50 per Million Tokens. Where It Beats Claude and GPT, Where It Trails, and the Region Setting EU Founders Must Check

TL;DR: What is Qwen4-Max?

Qwen4-Max is Alibaba's new flagship AI model, released on Monday 16 November 2026 after a five-day preview. It costs $1.50 per million input tokens and $4.50 per million output tokens in Alibaba's Singapore region, 25% less than Qwen3.8-Max, with a 1 million token context window. Alibaba's own charts put it first on 4 of 16 benchmarks and well behind Claude Opus 5.5 and Kimi K3.1 on others; Artificial Analysis ranks it 38th of 81 models. The weights are closed. For EU founders, the catch is the region: in Frankfurt it is only sold under a "Global" scope that can run your requests on servers in mainland China.
Qwen4 searches started climbing on Monday morning, a few hours after Alibaba published the Qwen4-Max launch post. Qwen is the model family many founders know from free, downloadable weights, so the first question everyone asked was simple: can I run it myself? This time the answer is no, and that changes how you should judge it.
Here is what Alibaba released, what it costs next to Claude and GPT, what its benchmark chart really says, and the region setting you need to check before a single customer record touches it.

What did Alibaba release on 16 November 2026?

Alibaba's Qwen team released Qwen4-Max (model ID qwen4-max), which QwenCloud calls the first model of the Qwen4 family. A preview version, qwen4-max-preview, opened on Alibaba's Token Plan subscription on 11 November and keeps its own preview discount until 25 November. Here is the short version of the spec sheet, from the launch post and the QwenCloud model docs:
  • Size: a sparse mixture-of-experts (MoE) model with 2.4 trillion parameters, of which 71 billion are active for each token. MoE means only a slice of the model works on each word, which keeps it cheaper to run.
  • Context window: 1 million tokens, roughly 750,000 English words, with text, images and video as input and text as output.
  • Output limits: up to 131,072 tokens of answer without thinking mode, and 262,144 tokens for thinking plus answer.
  • Tools: function calling, structured output, web search, a code interpreter, MCP (Model Context Protocol) and PDF reading.
  • Not supported: batch jobs, fine-tuning and private deployment, according to the docs.
  • Where to use it: the pay-as-you-go API, the Token Plan, Alibaba's QwenWork, Qoder and QwenStudio apps, and OpenRouter as qwen/qwen4-max.
  • APIs: compatible with the OpenAI Chat Completions and Responses formats and the Anthropic Messages format, so most agent code needs only a new endpoint address and model name.
Alibaba also says the model was trained with security checks on generated code, with offensive hacking data filtered out of training and extra safety data for chemical, biological, radiological and nuclear risks. Its internal review found "No capabilities posing large-scale cybersecurity risks", a claim no outside lab has tested yet. The South China Morning Post reports that Alibaba plans to scale its next generation, Qwen 5, to between 5 and 10 trillion parameters.

How much does Qwen4-Max cost?

The QwenCloud price list has one flat price for the whole 1 million token window, with no surcharge for long prompts and the same price in thinking and non-thinking mode:
Per 1 million tokensSingapore (International)Beijing and Global scope
Input$1.50$1.413
Output$4.50$4.24
Cached input (read)$0.075$0.071
Cache write, 5 minutes$1.875$1.767
Free tokens, pay-as-you-go1 millionNone listed
Source: QwenCloud price list and model docs, read 16 November 2026. The Global scope covers Frankfurt, Hong Kong, Tokyo and Virginia.
Against the previous flagship, Qwen3.8-Max at $2 and $6, input and output are 25% cheaper, and cached input drops 70%, from $0.25 to $0.075. That cache price matters most for agents, which resend the same long instructions and files on every step. Next to the Western models we covered recently, Qwen4-Max is a budget option: Claude Fable 5.2 lists at $8 and $40, GPT-6.1 Terra at $1 and $5, and Gemini 4 Hydrogen at $1.50 for input.
One warning on price. Artificial Analysis calls Qwen4-Max "very verbose": it used 127 million output tokens to finish its test suite, against a median of 78.6 million for comparable models, and its cost per test task came out at $3.27, rank 44 of 80. A cheap token that comes in large numbers is a smaller saving than the price list suggests, which is why we keep saying that AI token costs should be budgeted per task.

Is Qwen4-Max better than Claude, GPT and Kimi?

On some tasks, according to Alibaba. Its launch chart compares Qwen4-Max with Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Moonshot's Kimi K3.1 on 16 benchmarks. Read it row by row and Qwen4-Max has the top score on 4 of the 16:
BenchmarkQwen4-MaxBest rival on the chart
DeepSWE v1.1 (agentic coding)75.7 (top)Kimi K3.1, 75.5
Agents' Last Exam v1.146.7 (top)Opus 5.5 and Kimi K3.1, 43.4
HLAB-Vals 1.1 (legal work)21.3 (top)Kimi K3.1, 11.7
APEX-Agents v1.172.1Opus 5.5, 73.5
Terminal-Bench 4.048.5Kimi K3.1, 67.2
ProgramBench31.5Opus 5.5, 65.0
Humanity's Last Exam51.7Opus 5.5, 59.8
Source: Alibaba's benchmark chart in the launch post, 16 November 2026. These are Alibaba's own numbers and haven't been reproduced by outside labs.
The headline number in most reports, a 60% lower cost per task, comes from Work360, a test Alibaba built itself across 721 fields of office and expert work. The English launch post says the average task cost fell from $4.43 to $1.67, a 62% drop; the Chinese version gives 58% in yuan. Alibaba also says the model needs about 35% fewer tokens than Qwen3.8-Max at high effort on Work360. The first independent score is less flattering: Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index, rank 38 of 81, a little above the median of 41 for comparable models.
My reading: Qwen4-Max is a strong, cheap option for long agentic coding and document-heavy work, and it is behind the best Western models on terminal and from-scratch programming tasks. If you run coding agents in a terminal all day, Opus 5.5 or Kimi K3.1 still look stronger on this chart.

Can you download Qwen4-Max and run it yourself?

No. Qwen4-Max is a closed model, available only through Alibaba's cloud and partners such as OpenRouter. That is a break from habit: the previous flagship, Qwen3.8, has public weights on Hugging Face, and The Next Web notes that Alibaba is keeping the Qwen4-Max weights to itself. The launch post says "Smaller models in the family are coming to open source", with no sizes, licence or date. On 16 November the Qwen account on Hugging Face had no Qwen4 models at all.
If you picked Qwen because you could host it on your own servers, keep using the open Qwen3.8 models for now and wait for the smaller Qwen4 releases. There is one more reason the 2.4 trillion figure matters: China's draft AI export controls, which we covered in our piece on China's AI export controls, name models above 2 trillion parameters, while exempting API access and already published open weights.

Where does your data go when you use Qwen4-Max from Europe?

This is the part most launch coverage skipped. Alibaba sells Qwen4-Max in three ways, per the QwenCloud docs:
  • Singapore (International): its own price list and the 1 million free tokens.
  • Beijing (Chinese mainland).
  • Global scope, with endpoints in Frankfurt, Hong Kong, Tokyo and Virginia.
Alibaba's regions page says the Global scope runs inference on "Any available node, including the Chinese mainland and other regions". QwenCloud's regions page also describes separate EU-only and US-only scopes, and Qwen3.8-Max is listed under a US scope in Virginia. On 16 November, Qwen4-Max was listed under the Global scope only in all four of those cities.
So a Frankfurt endpoint doesn't mean your prompts stay in Frankfurt. If you process personal data of EU customers, that is a transfer question under the GDPR, with your privacy policy and your data processing agreements on the line. If your contracts promise EU-only processing, Qwen4-Max can't meet that promise today. It also connects to a story from October: Cohere reported that an open-weight Qwen model tried to send a test string to a Beijing cloud bucket during a sandboxed agent test, which we broke down in our article on the Qwen Beijing upload. Researchers mostly blamed messy training environments, and there is no sign of the same behaviour in Qwen4-Max, but it shows why you sandbox any agent before it gets real access.

What does Qwen4-Max mean for your startup?

If your AI bill is a real line in your budget, Qwen4-Max is worth a test this week. Here is how I'd run it:
  1. Decide the region before you write code. Singapore has separate pricing and the free tokens; the Global scope can process requests in mainland China. Write the choice into your vendor management notes.
  2. Test with your own tasks. Take 20 real jobs from last month (support replies, code reviews, contract summaries) and run them on Qwen4-Max and your current model. Alibaba's chart and Work360 are its own tests.
  3. Measure cost per finished task. Log tokens in and out for each job. A verbose model at $4.50 can cost more per task than a terse one at $5.
  4. Use the cache. At $0.075 per million cached input tokens, long system prompts and reference files are almost free to reuse. Put the fixed part of your prompt first so the cache can hit.
  5. Keep sensitive data out of the test. Use synthetic or public documents until your lawyer or DPO has looked at the Global scope.
  6. Plan around the gaps. No batch API and no fine-tuning means overnight bulk jobs and custom-trained variants stay on another provider.
  7. Build in a fallback. Because Qwen4-Max speaks the OpenAI and Anthropic API formats, switching back is a config change. Keep the model name in one config file, as in our guide to the startup tool stack.
For the month's other launches and price moves, see our AI news for November 2026 and the open-source AI news for November 2026. And if you're comparing this with the October price cuts, our recap of the AI price war has the baseline.
Are you getting customers from ChatGPT yet?

More and more traffic is coming from Perplexity, ChatGPT, Grok and other AI tools.

👉 Rank on ChatGPT

What we don't know about Qwen4-Max yet

  • Independent benchmark results beyond the first Artificial Analysis listing, and its place on the public Arena leaderboards.
  • When the smaller open-weight Qwen4 models arrive, in which sizes and under which licence.
  • Whether an EU-only or US-only scope will be added for Qwen4-Max, as Qwen3.8-Max already has in Virginia.
  • The exact output limit. The model page shows 393K tokens; the docs say 131,072 without thinking and 262,144 with it. We use the docs.

FAQ: Qwen4-Max

What is Qwen4-Max?

Qwen4-Max is Alibaba's flagship AI model, released on 16 November 2026. It is a 2.4 trillion parameter mixture-of-experts model with 71 billion active parameters, a 1 million token context window and image and video input.

How much does Qwen4-Max cost?

In Alibaba's Singapore region it costs $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075. In Beijing and the Global scope the prices are $1.413 and $4.24.

Is Qwen4-Max open source?

No. Qwen4-Max has closed weights and runs only through Alibaba's cloud and partners such as OpenRouter. Alibaba says smaller Qwen4 models will be released as open source, without a date.

Is Qwen4-Max better than Claude Opus 5.5?

On Alibaba's own chart, Qwen4-Max leads on 4 of 16 benchmarks, including DeepSWE agentic coding, and trails Opus 5.5 widely on ProgramBench and Humanity's Last Exam. Artificial Analysis ranks it 38th of 81 models.

Does Qwen4-Max keep EU data in Europe?

Not by default. In Frankfurt it is listed only under Alibaba's Global scope, which can run inference on any available node, including mainland China. No EU-only scope was listed for it on 16 November 2026.

How do I try Qwen4-Max?

Use the QwenCloud pay-as-you-go API, which accepts the OpenAI and Anthropic API formats, the Token Plan, Alibaba's QwenWork, Qoder and QwenStudio apps, or OpenRouter under the name qwen/qwen4-max.

About the Author

Violetta Bonenkamp, also known as MeanCEO, is an experienced startup founder with an impressive educational background including an MBA and four other higher education degrees. She has over 20 years of work experience across multiple countries, including 5 years as a solopreneur and serial entrepreneur. Throughout her startup experience she has applied for multiple startup grants at the EU level, in the Netherlands and Malta, and her startups received quite a few of those. She’s been living, studying and working in many countries around the globe and her extensive multicultural experience has influenced her immensely.
2026-11-16 10:26 News and Trends