AI News · 5 min read
Qwen4-Max preview is Token Plan only, and it ends Nov 25
Alibaba opened qwen4-max-preview on November 11 with a 1M context and no pay-as-you-go. The preview is retired November 25, and the key is not for app backends.
What Qwen shipped on November 11
Qwen4-Max is in preview, and the preview has an end date. On November 11, 2026, the official Qwen account posted ↗: "Qwen4 is coming." The same post says Qwen4-Max is in preview on Alibaba's Token Plan, with early access on QwenWork, Qoder, the Qwen app, and QwenStudio. Tom's Hardware Italia ↗ and PC Watch ↗ both covered that post the same morning.
These rows are from Alibaba Cloud's Qwen4-Max-Preview page ↗, last updated November 11. Thinking is on by default, at reasoning_effort xhigh unless you set low or medium.
| Spec | What the docs say |
|---|---|
| Model ID | qwen4-max-preview (bills show alias qwen4.0-max-preview) |
| Access | Token Plan only. No pay-as-you-go |
| Window | 10:00 Nov 11 to 24:00 Nov 25, 2026 (UTC+8), then retired |
| Context / output | 1M tokens / 393,216, shared with thinking |
| Input and tools | Text, image, video; function calling, search, JSON |
| APIs | OpenAI Chat Completions and Responses; Anthropic Messages |
Alibaba has not published a parameter count, a benchmark table, a model card, or downloadable weights. PC Watch and Tom's Hardware both flag that gap. This is not an open-weight release. For anything you can download, our open-source LLM comparison still starts the Qwen line at Qwen3.8-27B.
Who can call it, and what the key forbids
A normal Model Studio key does not open this model. 秋/十三 ↗ tried a regular Bailian key on November 11: qwen4-max-preview returned Access denied, and qwen4-max (without -preview) came back as a model that does not exist. That matches the docs.
Token Plan Personal uses a different key. The quick start ↗, also updated November 11, says those keys start with sk-sp- and cannot be swapped with a standard sk- key. Personal Edition is available only in the Singapore region. The OpenAI-compatible base URL is https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1. The Anthropic-compatible URL is https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic.
The personal plan page ↗ limits what that key is for. It is "intended for interactive use only within coding tools and agent tools such as Claude Code, Cursor, Qwen Code, Qoder and OpenClaw." It "must not be used for automation scripts, custom application backends, or any non-interactive batch call scenarios." Alibaba says a key used outside that scope can get the subscription suspended or the key banned. Inference on this plan runs in Global mode, so prompts and outputs cross borders. Compliance is on the subscriber.
Team Edition is a separate product. The team overview ↗ says it currently supports only the China (Beijing) region, at ¥150, ¥550, or ¥1,298 per seat each month. It has the same interactive-use rule. A Singapore personal key and a Beijing team key are different endpoints.
| Personal plan | Monthly price | Credits |
|---|---|---|
| Lite | $6, or $3 for a new user | 20,000 |
| Essential | $10, or $5 for a new user | 45,000 |
| Standard | $18 | 80,000 |
| Pro | $43 | 200,000 |
Those prices buy the plan. Alibaba has not published a per-token rate for Qwen4-Max. New-user prices last until 23:59:59 on November 30, 2026 (UTC+8), and only for accounts that have never bought Token Plan Personal. During the preview, credits for this model go "as low as 20% of the standard rate" (Alibaba's "5x usage"). From 22:00 to 08:00 UTC+8, a further 60% off takes that to 8% of the standard rate. The page never prints the standard rate. Unused monthly credits do not roll over, and a spent quota pauses the service.
What early users are reporting
Eric ↗ posted "Room 621," a 3D office he says Qwen4-Max built from one brief. That is his demo, not a score. Tyler White ↗ routed the model from Qoder's China app through a proxy into DeepSeek Harness and said a high-reasoning pelican test had run 50 minutes without finishing. One proxied run is not a latency number. It does match the docs: thinking is on by default and shares the output budget with the answer.
ThoufeekXLabs ↗ paired the preview with Claude Fable 5.2. Fable has a public token price. This preview does not. See our Fable 5.2 write-up.
Reddit was not reachable for this story.
We are not connecting this preview to Qoduck. Qoder's stealth model is still unidentified, and the November 11 post does not say Qoduck was a Qwen4 test. The earlier piece is here: Qoduck on Qoder.
Our take
Use the two weeks only if the work can disappear on November 25. The model ID is scheduled to come off Token Plan that night, Beijing time. Alibaba says the weights will not change during the window, which is what you want for a short bake-off and what you cannot build a December product on.
Keep the sk-sp- key inside a coding agent. The terms bar custom backends and batch jobs. If you need a normal API key and a price per million tokens, this is the wrong product. Compare it inside Claude Code, Cursor, or Qoder, then turn it off. Our best AI for coding guide is the map for tools that publish a rate.
Leave local setups on the current open weights. Nothing here is downloadable. Qwen3.8-27B and Arete are still the Qwen models you run yourself, and the open-source LLM ranking says the same. For a free coding agent with known terms, start with free AI coding tools.
Who should spend the credits this week: people who already live in a coding agent and want a look before the ID is retired, especially from 22:00 to 08:00 UTC+8, when the credit rate drops to 8% of standard. Who should skip it: anyone pricing a production agent, anyone who wanted weights, and anyone whose Bailian key returned Access denied and who does not want another subscription.