Qwen4-Max Preview: Specs, Token Plan Price, Claude Code Setup and Deadline

The first Qwen4 model is here, but only for two weeks. Alibaba switched on Qwen4-Max-Preview on 11 November 2026, and it will be retired at the end of 25 November.

It has a 1M-token context, reads images and video, and works inside Claude Code, Qwen Code, Cursor and other coding agents. The catch: it is only sold through Alibaba’s Token Plan, a subscription from $3 for new users.

This guide covers what is confirmed, what is still unknown, the plan prices, step-by-step setup for Claude Code and Qwen Code, the settings that trip people up, and how it compares with other big models right now.

What Qwen4-Max-Preview Is

Qwen4-Max is the top model of Alibaba’s new Qwen4 family. Alibaba listed Qwen4-Max, Qwen4-Flash, Qwen4-Plus and Qwen4-27B as upcoming at its Apsara Conference in Amsterdam on 29 October 2026.

The Max model arrived first as a managed preview. Alibaba describes it as its next-generation multimodal flagship, aimed at coding, office work, front-end builds and long agent tasks with many tool calls.

The Qwen team says it improves on Qwen3.8 in world knowledge, code generation and staying on track over long tasks. Those are vendor claims; no benchmark table came with the launch.

Note: the discount deadline (30 Nov) is later than the model’s retirement (25 Nov). Do not buy a plan on 28 November expecting to use Qwen4-Max.

Confirmed Specs and What Is Still Unknown

Image and video limits are from a developer write-up on DEV Community that summarises Alibaba’s docs. Everything else comes from Alibaba’s Qwen Cloud documentation.

Not published yet

• Benchmarks: none. No model card either.
• Parameter count and architecture: not disclosed.
• Per-token price: not published, because pay-as-you-go is not offered.
• Final release: no date for a stable Qwen4-Max, and no word on whether it will match the preview.
• Open weights: AlphaSignal suggests a smaller Qwen4-27B is the likelier open release. That is analysis, not an Alibaba statement.

Token Plan Prices and Credits

The Token Plan is a monthly subscription that gives you credits to spend on Qwen and some third-party models inside coding and agent tools. Qwen4-Max-Preview is only available this way.

Prices are from Qwen Cloud’s pricing page and two independent write-ups that checked it. The final price is set at checkout, so confirm it there.

• Preview discount: Qwen4-Max-Preview uses credits at as little as 20% of the standard rate.
• Overnight discount: calls between 22:00 and 08:00 UTC+8 use even fewer credits (write-ups quote 4-8% of standard; your console Analytics page shows the real rate).
• No rollover: unused credits expire at the end of each month.
• No refund on Individual: you cannot unsubscribe mid-term; you can only stop renewal.
• Team Edition: separate seats (Standard 25,000, Pro 100,000, Max 250,000 credits) with a no-training commitment.

Tip: Alibaba has not published how many credits one Qwen4-Max call costs. Run a few typical tasks on day one and check Analytics before planning a big job around it.

For a wider view of paid API costs, see my AI API pricing comparison.

How to Try It Without Paying

If you only want to see how it answers, you do not need the Token Plan.

1. Open Qwen Chat. Go to chat.qwen.ai in your browser.
2. Pick the model. Choose Qwen4-Max-Preview from the model list.
3. Sign in when asked. A Qwen account is required for this model.

AlphaSignal also lists access through Alibaba’s QwenWork, Qoder, Qwen App and QwenStudio. Usage limits in those apps were not published.

Set It Up in Claude Code

Alibaba publishes an official Claude Code config for the Token Plan. It maps Claude Code’s “Opus” slot to Qwen4-Max-Preview and uses smaller Qwen models for quick tasks.

1. Subscribe to a Token Plan. Set your Qwen Cloud console region to Singapore, or the Individual plan may not appear.
2. Copy the Token Plan key. It is separate from any pay-as-you-go key. Token Plan keys start with sk-sp-.
3. Edit your settings file. Open ~/.claude/settings.json on macOS/Linux or C:\Users\<username>\.claude\settings.json on Windows.
4. Paste the config below and replace YOUR_API_KEY.
5. Skip the Anthropic login. In ~/.claude.json, set "hasCompletedOnboarding": true.
6. Open a new terminal and test. Run claude -p "Hello", then /status inside Claude Code to confirm the base URL.

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "YOUR_API_KEY",
    "ANTHROPIC_BASE_URL": "https://token-plan.maas.qwencloudapi.com/apps/anthropic",
    "ANTHROPIC_MODEL": "auto",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3.6-flash",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3.8-flash",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen4-max-preview",
    "CLAUDE_CODE_SUBAGENT_MODEL": "auto",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "983616"
  }
}

• Force Qwen4-Max everywhere: change ANTHROPIC_MODEL from auto to qwen4-max-preview. “Auto” lets Alibaba pick a model per request.
• Long answers cut off: Alibaba recommends adding "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000" for this model.
• Full 1M context: set CLAUDE_CODE_MAX_CONTEXT_TOKENS to 1000000.

Want to switch between several providers inside Claude Code? My Claude Code Router guide shows how to keep profiles for each one.

Set It Up in Qwen Code and Other Tools

Qwen Code (desktop)

1. Open model settings. Go to More, then Settings, then Models, then Add Model.
2. Choose the provider. Type QwenCloud, Provider Token Plan, Region Singapore, and paste your key.
3. Select the model. Pick qwen4-max-preview from the list and save. Avoid “Auto” if you want Qwen4-Max every time.

Cline, Cursor, OpenCode, Codex

Tools that ask for a base URL need the endpoint that matches their protocol:

Set the model name to qwen4-max-preview and use your Token Plan key. If your console shows a different URL, use the console’s.

Quick connection test with curl

curl https://token-plan.maas.qwencloudapi.com/compatible-mode/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_PLAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen4-max-preview",
    "messages": [{"role": "user", "content": "Say hello in one line."}],
    "max_tokens": 8192,
    "reasoning_effort": "low"
  }'

API Settings That Matter

Qwen4-Max behaves a little differently from older Qwen models. These four settings cause most of the surprises.

With the official OpenAI Python SDK, pass reasoning_effort and enable_thinking through extra_body. On the Responses API, the setting is reasoning.effort.

Tip: for everyday edits, medium is a sensible start. Save xhigh for hard bugs and planning. My reasoning effort guide explains the speed and cost trade-off across models.

Qwen4-Max vs Qwen3.8-Max, LongCat-2.5 and Fable 5.2

There are no Qwen4-Max benchmarks to compare yet, so this table compares access, price and limits, which you can act on today.

For reference, Qwen3.8-Max scored 86.6 on Terminal Bench 2.1 and 67.7 on SWE-bench Pro (vendor-reported, via AlphaSignal). Qwen4-Max should beat those if Alibaba’s claims hold, but nobody has shown numbers yet.

• Try Qwen4-Max if you want a cheap look at Alibaba’s next flagship inside your coding agent this month.
• Stay on Qwen3.8-Max for anything you need next month; the preview disappears on 25 November.
• Pick LongCat-2.5 if you want open weights or plain per-token API billing. Details in my LongCat-2.5 guide.
• Run Qwen locally instead with Qwen3.8-Flash-Next, using my local Qwen3.8-Flash-Next guide.

Usage Rules, Privacy and Limits

Warning: the Individual Token Plan is for interactive use in coding and agent tools only. Alibaba’s FAQ says automation scripts and app backends are not allowed, and breaking that can get your key revoked.

• No key sharing: keys are personal. Public exposure can trigger automatic suspension.
• Region: the Individual plan is sold for Singapore with global inference, so prompts may cross borders. Choosing Singapore does not keep processing there.
• Training: Alibaba’s Token Plan FAQ says conversation data is not used for model training.
• Rate limits: counted per account. Short bursts can hit a 429 even under the per-minute total.
• One subscription per account: you can hold one Individual and one Team plan, but keys are not interchangeable.
• Work data: only paste code your employer allows into a preview model hosted abroad.

Troubleshooting

My Take

A two-week, subscription-only preview with no benchmarks is an unusual way to launch a flagship. It feels like Alibaba wants real agent traffic from developers before it locks in the final model and its price.

For readers, the deal is decent: $3 to $6 buys a month of credits, and the preview burns them at a fraction of the normal rate. If you already use Claude Code or Qwen Code, it costs very little to point the Opus slot at Qwen4-Max and judge it on your own repo.

I would not build anything that depends on it. The model vanishes on 25 November, nobody has published scores, and the plan rules forbid backend use.

Treat it as a free look at where Qwen4 is heading, and watch for Qwen4-27B if you care about open weights. Kimi fans can compare it with my Kimi K3.1 Pro guide.

FAQ

Is Qwen4-Max open source?

No. The preview is proprietary and API-only, with no weights or licence for self-hosting announced. A smaller Qwen4-27B is expected but its licence is not confirmed.

When does the Qwen4-Max preview end?

At 24:00 on 25 November 2026 (UTC+8), which is 16:00 UTC that day. Alibaba says the model will be retired after that.

How much does Qwen4-Max cost?

There is no per-token price. You need a Token Plan: Individual plans cost $6, $10, $18 or $43 a month, with $3 and $5 new-user offers on the two cheapest tiers until 30 November.

Can I use Qwen4-Max for free?

You can chat with it at chat.qwen.ai after signing in. API access for coding tools needs a paid Token Plan, which has no free trial.

Does Qwen4-Max work with Claude Code?

Yes. Point ANTHROPIC_BASE_URL at https://token-plan.maas.qwencloudapi.com/apps/anthropic, use your Token Plan key, and set the Opus model (or ANTHROPIC_MODEL) to qwen4-max-preview.

What context window does Qwen4-Max have?

1M tokens of input, with up to 393,216 output tokens shared between thinking and the answer.

Are there Qwen4-Max benchmarks?

Not yet. Alibaba published no benchmark table or model card with the preview, so treat any scores you see as unofficial until Alibaba or an independent lab releases them.

Can I use a Token Plan key in my own app?

Not on the Individual plan. Its rules allow interactive use in coding and agent tools only, not automation scripts or app backends.

Sources

• Qwen Cloud docs: Latest model, Qwen4-Max-Preview (docs.qwencloud.com)
• Qwen Cloud docs: Token Plan, Claude Code setup and FAQ
• Qwen Cloud Token Plan pricing page (qwencloud.com/pricing/token-plan)
• @Alibaba_Qwen announcement on X, 11 November 2026
• AlphaSignal: “Alibaba Launches Qwen4-Max With Agent-First Coding and 1M Token Support”
• ai-tldr.dev: Qwen4-Max-Preview model page
• DEV Community: “Qwen4-Max Preview: Token Plan Only, Default Is xhigh, and It Retires on Nov 25”
• note.com (kazu): Qwen4-Max Token Plan setup and pricing notes