Alibaba Cloud opened a limited preview of Qwen4-Max-Preview on November 11, 2026. Alibaba’s Model Studio docs describe it as the next-generation flagship multimodal model in the Qwen family, built for coding, office productivity, frontend design and complex agentic tasks. It takes text, images and video, has a 1M-token context window, and speaks the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs.
Access is the catch. The model is available through Token Plan only, Alibaba’s Credits-based subscription, and pay-as-you-go isn’t supported. The preview runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8), which is 16:00 UTC on November 25. Alibaba says the model won’t change during the preview and will be retired when it ends.
That makes it a two-week test window, not a production model. Because it accepts Anthropic Messages requests, you can point Claude Code at it in a few minutes, and Alibaba’s own Claude Code guide already maps it as the Opus-tier model.
What Alibaba lists for qwen4-max-preview
Alibaba’s model page gives these specs:
qwen4-max-preview at a glance
| Model ID | qwen4-max-preview (alias qwen4.0-max-preview; bills show qwen4.0-max-preview) |
|---|---|
| Input | Text, images and video |
| Context length | 1M tokens |
| Maximum output | 393,216 tokens, shared by thinking and the final answer |
| Thinking | On by default; reasoning_effort low, medium or xhigh (default xhigh) |
| Tools | Function calling, web search and built-in tools |
| Structured outputs | JSON Object and JSON Schema |
| API compatibility | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages |
| Access | Token Plan (Personal or Team) only; no pay-as-you-go |
| Preview window | Nov 11, 10:00 to Nov 25, 24:00, 2026 (UTC+8), then retired |
Alibaba’s model page doesn’t list any benchmark results for Qwen4-Max-Preview. If you want to know how it compares with Claude, GPT or Gemini models on your code, the preview window is the time to run your own tasks against it.
Token Plan: the only way in
Token Plan is Alibaba Cloud Model Studio’s subscription for AI coding and agent tools, billed in Credits. It comes in Personal and Team editions, and Token Plan Personal is available only in the Singapore region. Alibaba’s Personal Edition page lists four tiers:
| Plan | Price | Monthly quota | Concurrent agents |
|---|---|---|---|
| Lite | $6/month ($3/month new-user price) | 20,000 Credits | 1-2 |
| Essential | $10/month ($5/month new-user price) | 45,000 Credits | 2-3 |
| Standard | $18/month | 80,000 Credits | 3-4 |
| Pro | $43/month | 200,000 Credits | 6-8 |
The new-user prices run until 23:59:59 on November 30, 2026 (UTC+8) and apply only to accounts that have never bought a Token Plan Personal subscription. Extra bundles cost $15 per bundle per month for 20,000 Credits, need an active subscription, and you can hold up to 5. When the monthly quota and any bundle Credits run out, the service pauses until the next subscription month.
Alibaba also lists a preview promotion. Credits consumption for qwen4-max-preview is as low as 20% of the standard rate. Calls between 22:00 and 08:00 (UTC+8) get an extra 60% off, which Alibaba puts at 8% of the standard rate. Alibaba doesn’t publish a fixed Credits-per-token figure. The page says the Credits per request depend on model type, token usage, thinking mode and tool calls, and points you to the usage details in the console.
Two more terms matter before you subscribe. Token Plan runs in the Singapore region with Global deployment, so Alibaba says inference happens globally and prompts and outputs involve cross-border data transfer. And the Personal Edition doesn’t currently support unsubscribing.
Pointing Claude Code at qwen4-max-preview
Alibaba’s Claude Code guide uses the Anthropic-compatible Token Plan endpoint. For the CLI, it sets these environment variables, with YOUR_API_KEY replaced by your Token Plan API key:
export ANTHROPIC_AUTH_TOKEN="YOUR_API_KEY"
export ANTHROPIC_BASE_URL="https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic"
export ANTHROPIC_MODEL="auto"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen3.6-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3.8-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen4-max-preview"
export CLAUDE_CODE_SUBAGENT_MODEL="auto"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="983616"With this mapping, the main model is auto, Alibaba’s platform model that picks an underlying model per request. qwen4-max-preview only answers when Claude Code asks for the Opus tier. To run the main session on the preview model, you can set ANTHROPIC_MODEL to qwen4-max-preview. That’s a change to Alibaba’s example, not part of it, and subagents still follow CLAUDE_CODE_SUBAGENT_MODEL.
These variables only last for the current terminal. For a persistent setup, the same guide puts them in the env block of ~/.claude/settings.json (C:\Users\<username>\.claude\settings.json on Windows):
{
"env": {
"ANTHROPIC_AUTH_TOKEN": "YOUR_API_KEY",
"ANTHROPIC_BASE_URL": "https://token-plan.ap-southeast-1.maas.aliyuncs.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"
}
}Then check the connection:
claude -p "Hello"Alibaba adds one note for this model: if long output gets truncated with qwen4-max-preview, set CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000. For more on how Anthropic-compatible endpoints plug into Claude Code, see our guide to Anthropic-compatible endpoints.
Token Plan isn’t the same product as Alibaba’s Coding Plan, which uses a different endpoint (coding-intl.dashscope.aliyuncs.com) and, in the same guide, maps Claude Code to qwen3.7-plus. A Coding Plan key won’t reach the preview. We compared Coding Plan with pay-as-you-go billing in Alibaba Coding Plan vs pay-as-you-go.
The max_tokens change that can cut answers short
Alibaba flagged a behaviour change in its API compatibility notice dated November 2, 2026. Starting with qwen4-max-preview, max_tokens on the OpenAI Chat Completions API and the DashScope API is equivalent to max_completion_tokens. Both cap thinking and the final answer combined.
Thinking is on by default, at xhigh unless you change it, and it draws from the same budget as the answer. A max_tokens value you sized for the answer alone on an older Qwen model now has to cover the reasoning too. The model can spend most of that budget thinking and return a cut-off answer. Alibaba says existing Qwen models keep their previous behaviour, which is why this shows up only when you swap the model ID.
Before you test qwen4-max-preview through an OpenAI-compatible client
- Raise max_tokens (or max_completion_tokens) to leave room for thinking plus the answer
- Pick reasoning_effort low, medium or xhigh; don't set reasoning_effort and thinking_budget together
- In multi-turn Chat Completions, return earlier thinking in reasoning_content, not inside content
- Use the Token Plan API key and the token-plan.ap-southeast-1.maas.aliyuncs.com base URL
- Keep calls interactive inside a coding or agent tool, per the Token Plan usage terms
- Plan to switch models before 24:00 on November 25, 2026 (UTC+8)
The multi-turn point comes from the same model page. Chat Completions enables preserve_thinking by default for this model, and Alibaba asks you to send historical thinking back in the assistant messages’ reasoning_content field rather than concatenating it into content.
Other tools on the same plan
Alibaba’s Token Plan overview lists configuration guides for Qoder, Claude Code, Codex, Qwen Code, OpenClaw, Hermes Agent, OpenCode, QwenWork, Trae, DeepSeek Harness and WorkBuddy, plus a generic integration guide. The overview also offers a Bailian CLI path that signs in with bl auth login --config token-plan against the same base URL. If you already run Qwen models in another client, our earlier Qwen3.7-Max pricing and setup notes cover a pay-as-you-go Qwen option.
Is the Qwen4-Max-Preview window worth using?
If you already pay for Token Plan, yes: switch the Opus mapping or the main model to qwen4-max-preview, run it on real tasks from your own repos, and watch the Credits it burns in the console. The preview promotion and the night discount make this the cheapest time to try it, and Alibaba has said the model won’t change before November 25.
If you don’t have Token Plan, weigh the subscription against a model you can’t keep. The preview ends on November 25, 2026, the Personal Edition doesn’t support unsubscribing, and the plan’s terms rule out scripts and backends. A Lite plan at the new-user price is a low-cost way to look, but don’t plan any pipeline around qwen4-max-preview itself.
