News

Qwen4-Max: Price, Benchmarks and Who Should Try It

Alibaba's Qwen4-Max is out with a 1M-token context and $1.50/$4.50 per million tokens. What the benchmarks really show, and who should test it now.

Alibaba on Monday, November 16, made Qwen4-Max generally available. It’s the first model in the Qwen4 family and the company’s new flagship, five days after a limited preview opened on November 11. The pitch is simple: near-frontier coding and agent performance, a 1-million-token context window, and a list price of $1.50 per million input tokens and $4.50 per million output tokens.

That price is the real story. The benchmark chart is mostly Alibaba grading its own homework, but the one independent score out so far puts Qwen4-Max in the same neighborhood as models that cost several times more. If you pay for tokens at volume, this release is worth an afternoon of testing. Here’s what launched, what the numbers really say, and who should try it.

What Alibaba launched

According to Alibaba Cloud’s Model Studio release notes, Qwen4-Max is “the next-generation flagship multimodal model in the Qwen family,” built for coding, office work, front-end design and complex agentic tasks. It takes text, image and video input, returns text, and supports web search, tool calling and structured outputs. It’s also compatible with both the OpenAI and Anthropic APIs, so most existing client code can point at it by swapping a base URL and a model name.

The QwenCloud model page lists the specs that matter for builders:

  • Context: 1 million tokens, with up to 393K output tokens and up to 262K reasoning tokens.
  • Price: $1.50 per million input tokens and $4.50 per million output tokens on the international (Singapore) endpoint. Cached input reads drop to $0.075 per million.
  • Batch: QwenCloud’s pricing docs say Batch API calls cost 50% of real-time rates. Thinking tokens bill as output, and thinking and non-thinking modes cost the same.
  • Regions: the same docs list a global deployment that includes US (Virginia), alongside Frankfurt, Hong Kong and Tokyo.

Under the hood, IT Home reports, citing the Qwen team’s launch blog, that Qwen4-Max is a sparse mixture-of-experts model with 2.4 trillion parameters. Three of every four layers use Gated DeltaNet, which compresses history into a fixed state, and the fourth uses a new block-level “Qwen Sparse Attention.” That design is aimed squarely at long contexts: it keeps attention costs from exploding as inputs grow.

It’s not open-weight. Alibaba didn’t release downloadable weights for Qwen4-Max, though it said smaller Qwen4 models will be open-sourced later, as Bloomberg reported via The Straits Times. That matters, because Qwen’s open models are the most downloaded open-weight family in the world, and many teams know the brand from running it locally.

The benchmarks, read honestly

Alibaba told Bloomberg that Qwen4-Max scored near the top on coding benchmarks and beat Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro. Treat that as a claim, not a verdict. An Apidog breakdown of the launch chart counts 16 benchmarks, of which 10 are Alibaba’s own runs. Qwen4-Max leads on 4 of the 16 and trails clearly on others, including ProgramBench (31.5 vs. 65.0) and Terminal-Bench 4.0, where it scores 48.5 while the four non-Qwen models on the chart land between 56.1 and 67.2.

The independent check is Vals AI. On the Vals Index, Qwen4-Max scores 63.20%, ranking 11th of 38 models and first among Chinese models, just ahead of Kimi K3.1 Pro at 62.52%. For context, the top of that index is Claude Fable 5.2, which we covered in our Claude Fable 5.2 breakdown, at 71.31%. Vals also measured 48.48% on Terminal-Bench 4.0, which matches Alibaba’s own number, and 86.66% on Vibe Code Bench.

So the fair summary is this: Qwen4-Max is a clear step up from Qwen3.8-Max, a solid second-tier frontier model on independent tests, and not the best coding agent you can buy.

Why it matters anyway: cost per finished task

The more interesting claim is about efficiency. Alibaba says Qwen4-Max scored 52.69 on Work360, its own benchmark of professional tasks across 26 domains, up from 47.74 for Qwen3.8-Max, while the average cost per task fell from $4.43 to $1.67, as Runtime Wire and IT Home both report. Alibaba also says token use dropped by about 24% to 35% on Work360 and about 48% to 49% on Terminal-Bench 4.0, depending on reasoning effort.

That’s the right metric to watch. Price per token is only half the bill. A model that needs half the tokens to finish the same job can beat a cheaper-per-token rival on your invoice. If those savings hold up on your workloads, Qwen4-Max becomes a strong candidate for high-volume, long-context jobs: document review, data extraction, internal research agents and bulk code changes.

One caution from the launch data: on Alibaba’s own secure-coding test, the share of function-level code that passed both functional and security checks rose from 22.0% to 31.8%. Better, yes. But about two-thirds of outputs still failed, so generated code still needs review before it ships.

Who should test it, and who should wait

Test it now if you run agents or batch jobs where token spend is a top-three line item, you need very long context, or you already use the OpenAI or Anthropic SDKs and can switch models with a config change.

Wait if your main workload is long-running terminal or coding agents, where independent scores still favor the leading US models, or you need weights you can host yourself. The open Qwen4 models aren’t out yet.

Check with your security and legal teams first if you handle regulated or customer data. Qwen4-Max is a hosted service from a China-headquartered company. Many US companies are fine with that for some workloads and not others. Get a clear answer before anything sensitive goes through it, and follow the basics in our guide to using AI at work without leaking data.

How to evaluate it in an afternoon

  1. Pick 20 to 50 real tasks from your own logs, not public benchmark prompts. Include a few that your current model gets wrong.
  2. Run them on both models with the same prompts, tools and effort settings, and record the token counts from each API response.
  3. Score the outputs blind if you can, then compute cost per correct answer, not cost per token.
  4. Repeat twice. Model outputs vary run to run, and small gaps often disappear.

If you need a framework for this, our guides on how to choose an AI model and how to test an AI feature walk through the setup.

The bigger picture

Qwen4-Max lands the same week Nvidia reports earnings and Anthropic is set to start trading publicly, and it adds to a run of fast Chinese releases from Moonshot, Tencent and DeepSeek. The pattern is getting clear. The very top of the leaderboard still belongs to a few US labs, but the gap to “good enough at a fraction of the price” keeps shrinking. For most teams, that’s good news. It means the default model choice should be re-tested every quarter, not set once and forgotten.

My take: don’t switch on the strength of a launch chart. But if your AI bill is growing faster than your revenue, Qwen4-Max is one of the most credible cheaper options to put through a real test this month.

Kabeer Ahmed

Kabeer Ahmed is the founder and editor of The AI Year, where he explains AI tools, models, agents and AI security in plain English. He works on AI automations for businesses and is the founder of Runnex Logistics and of Aurel Digitals, a web design and automation agency. He writes practical guides for people using AI at work, covers selected news with links to primary sources, and reviews every post before it's published. Reach him at hello@theaiyear.com or on LinkedIn: https://www.linkedin.com/in/kabeerahmed09

Keep reading