Free guide

Qwen4-Max: Why Alibaba's New Model Costs a Third as Much per Job

Alibaba's full Qwen4-Max release scores higher than the model before it and, in Artificial Analysis' test, costs about a third as much per task. Here's where the savings come from, how to try it today, and where it still trails.

By Beau Johnson·
The Short this guide goes with: Qwen4-Max: Why Alibaba's New Model Costs a Third as Much per Job

Quick answer

Qwen4-Max is Alibaba's new flagship model, released November 16, 2026 (UTC), at $1.50 per million input tokens and $4.50 per million output tokens. Artificial Analysis measured it at $3.27 per Intelligence Index task versus $10.02 for Qwen3.8-Max while scoring 9 points higher. Most of that bill is the model reading, not writing, and rereading cached input now costs 70% less. You can call it on QwenCloud (model qwen4-max) or OpenRouter (qwen/qwen4-max).

The short version

  • Qwen4-Max is live on QwenCloud's API and on OpenRouter at $1.50 in and $4.50 out per million tokens, with a 1M-token context window.
  • Artificial Analysis scored it 43 on its Intelligence Index (Qwen3.8-Max: 34) at $3.27 per task (Qwen3.8-Max: $10.02). Most of that cost is the model reading, and cache hits are now 70% cheaper.
  • It is not the top model on that index: Kimi K3.1 Pro scores 46, and Qwen4-Max trails it on Terminal-Bench 4.0 (48% vs 67%).

How it actually works

Think of a letter writer who charges for every word he reads and writes

AI models bill by the token, a small chunk of text, both for what they read (input) and what they write (output). Your cost for one finished job is the rate times every word read and written. Qwen4-Max lowered the rate and changed how much gets read and written.

The rate dropped by a quarter, and the letters got shorter

Qwen3.8-Max cost $2.00 per million input tokens and $6.00 per million output tokens. Qwen4-Max costs $1.50 and $4.50, 25% lower on both. Artificial Analysis also counted about 105,000 output tokens per task, against about 171,000 for Qwen3.8-Max: 39% fewer.

Most of the bill is reading, not writing

Using Artificial Analysis' own numbers, writing was only about $1.03 of Qwen3.8-Max's $10.02 per task, and about $0.47 of Qwen4-Max's $3.27. The rest is reading, which fell from roughly $9.00 to roughly $2.80 per task. So the shorter answers help, but they are the smaller part of the drop.

Agents hand over the same pile every step

An agent sends the model the same system prompt, tool definitions and reference files again and again. QwenCloud reuses that shared start of the prompt from cache, and a cache hit now costs $0.075 per million tokens, down from $0.25: 70% less. That cheaper rereading is where most of the saving comes from. Alibaba's own Work360 test points the same way, with average cost per task falling from $4.43 to $1.67 in its internal evaluation.

How to set it up

  1. 1.Try it in Alibaba's apps

    Alibaba says you can try Qwen4-Max in Qwen Studio, QwenWork and Qoder. Pick Qwen4-Max in the model list and give it a task you already know the answer to.

  2. 2.Get a QwenCloud API key (free tokens to start)

    Create a key at home.qwencloud.com/api-keys in the Singapore region. QwenCloud's docs say pay-as-you-go calls in Singapore come with 1 million free tokens for this model, usually valid for 90 days; the free quota only counts with a Singapore key on the Singapore endpoint. Then call model qwen4-max on the OpenAI-compatible endpoint.

    export DASHSCOPE_API_KEY="YOUR_QWENCLOUD_API_KEY"
    curl https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1/chat/completions \
      -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"model": "qwen4-max", "messages": [{"role": "user", "content": "Plan a 3-step launch for my side project"}]}'
  3. 3.Or call it through OpenRouter

    If you already use OpenRouter, the model id is qwen/qwen4-max at the same $1.50 / $4.50 per million tokens (one provider, Alibaba Cloud).

    curl https://openrouter.ai/api/v1/chat/completions \
      -H "Authorization: Bearer $OPENROUTER_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"model": "qwen/qwen4-max", "messages": [{"role": "user", "content": "Hello"}]}'
  4. 4.Use it inside Claude Code

    QwenCloud's Claude Code page gives a pay-as-you-go setup on its Anthropic-compatible endpoint and says to set the Opus slot to qwen4-max and keep CLAUDE_CODE_MAX_CONTEXT_TOKENS at 983616. Put these in one terminal (or in the env section of ~/.claude/settings.json), then run claude and choose Opus with /model. If long answers get cut off, also set CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000.

    export ANTHROPIC_AUTH_TOKEN="YOUR_QWENCLOUD_API_KEY"
    export ANTHROPIC_BASE_URL="https://ap-southeast-1.maas.qwencloudapi.com/apps/anthropic"
    export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen4-max"
    export CLAUDE_CODE_MAX_CONTEXT_TOKENS="983616"
    claude -p "Hello"
  5. 5.Measure cost per finished job, not price per token

    Run the same real task on the model you use now and on Qwen4-Max, then compare what each run cost in total (OpenRouter's activity page or your QwenCloud bill shows it). A cheaper token only helps if the job gets done in fewer of them.

The fine print

  • Prices are Alibaba's list prices on November 16, 2026: $1.50 input, $4.50 output and $0.075 for cache hits per million tokens. Explicit cache writes cost extra ($1.875 for 5 minutes, $3 for 1 hour).
  • The $3.27 vs $10.02 figure is Artificial Analysis' cost to run its Intelligence Index, not a promise about your own workload. Alibaba's $4.43 to $1.67 figure is from its own internal test.
  • Qwen4-Max is not open weights. Alibaba says smaller Qwen4 models will get open weights later.
  • It trails Kimi K3.1 Pro on Artificial Analysis' Intelligence Index (43 vs 46) and on Terminal-Bench 4.0 (48% vs 67%).
  • API keys are region-specific, and the free tokens only apply in the Singapore region. Calls in other regions are billed from the first token.

FAQ

How much does Qwen4-Max cost?

$1.50 per million input tokens, $4.50 per million output tokens and $0.075 per million cached input tokens on QwenCloud and OpenRouter. That is 25% lower than Qwen3.8-Max on input and output, and 70% lower on cache hits.

Is Qwen4-Max cheaper than Kimi K3.1 Pro?

Per task, yes in Artificial Analysis' test: $3.27 per Intelligence Index task vs $3.68 for Kimi K3.1 Pro, about 11% less. Kimi K3.1 Pro still scores 3 points higher on that index.

Is Qwen4-Max free?

Not in general. QwenCloud's docs say pay-as-you-go calls in the Singapore region come with 1 million free tokens for the model, usually for 90 days. After that you pay per token.

Is Qwen4-Max open source?

No. Qwen4-Max is available through Alibaba's API and partners like OpenRouter. Alibaba says the open weights of smaller Qwen4 models will be released later.

Can I use Qwen4-Max in Claude Code?

Yes. QwenCloud documents an Anthropic-compatible endpoint for Claude Code. Set ANTHROPIC_BASE_URL to the QwenCloud endpoint, use your QwenCloud API key, and map the Opus slot to qwen4-max.

Sources

Checked November 15, 2026 (PT) against Alibaba's launch post and blog, the QwenCloud model page and docs (latest model, Claude Code, free quota), OpenRouter's listing and API, and Artificial Analysis' evaluation post.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides