Alibaba’s Qwen team opened its next flagship model on Tuesday night with an unusual condition attached: it switches off in two weeks. “Qwen4 is coming,” the team posted on X at 9:01 p.m. Eastern on November 10, and Qwen4-Max went into preview at the same time. Alibaba Cloud’s model page says the preview runs until midnight Beijing time on November 25 and that “the model will be retired after the preview ends.”
Qwen4-Max reads up to a million tokens, takes text, images and video, and is pitched at coding and agent work. What it doesn’t come with says more: Alibaba hasn’t published a benchmark score, a parameter count, a per-token price or a word about open weights. The only developer route is a monthly Token Plan subscription that starts at $3 for new customers and forbids the thing most developers buy an API for, which is wiring the model into their own app. Here’s what Qwen4-Max actually offers, what it costs, and the time-zone quirk that makes it cheapest during the American workday.
- Price: there’s no per-token price. Developers get Qwen4-Max only through Alibaba’s Token Plan, from $6 a month ($3 for new customers until November 30) up to $43, and preview calls use as little as 20% of the normal credit rate.
- Availability: November 10 at 9 p.m. ET to November 25 at 11 a.m. ET, then retired. Alibaba says early access is also open in the Qwen app, QwenWork, Qoder and QwenStudio.
- Specs: a 1M-token context window and up to 393,216 output tokens, shared between thinking and the answer. Text, images and video in, text out.
- The catch: Token Plan is limited to interactive use in coding tools such as Claude Code and Cursor (no app backends, scripts or batch jobs), the Individual plan can’t be cancelled, and Alibaba hasn’t shown a single benchmark.
Qwen4-Max specs: a million tokens in, 393,216 out
Alibaba calls Qwen4-Max-Preview “the next-generation flagship multimodal model in the Qwen family.” Its model ID is qwen4-max-preview, with an alias, qwen4.0-max-preview, that shows up on bills. According to the model page, it accepts text, images and video, supports function calling, web search and JSON Schema outputs, and works with OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages API. That last point matters more than it sounds: it’s what lets Qwen4-Max slot into Claude Code, Cursor and similar tools that already speak those formats.
Thinking is on by default at the highest setting, xhigh, with low and medium as the alternatives. The output limit is 393,216 tokens, but Alibaba is careful to say that number is “shared by thinking and the final answer.” It also flags a breaking change: starting with this model, max_tokens in Chat Completions caps thinking and answer together, so a limit sized for the visible answer alone can cut a response off. Older Qwen models keep the old behavior.
The Qwen team’s own pitch, in its launch post, is “deeper knowledge,” “richer world understanding” and “stronger coding and agentic capabilities.” Those are claims with no numbers behind them yet. Neither the model page nor the Token Plan documentation contains a benchmark table or a parameter count, and we found no Qwen4-Max page on Artificial Analysis when we checked Wednesday morning. Interest isn’t the problem: X’s counter showed the launch post past 860,000 views by then.
No per-token price: Qwen4-Max is rented by the month, in Credits
The model page is blunt about access: Qwen4-Max-Preview “is available through Token Plan only. Pay-as-you-go access is not supported.” It isn’t on QwenCloud’s model price list at all. For comparison, Alibaba’s current pay-as-you-go flagship, Qwen3.8-Max, lists there at $2 per million input tokens and $6 per million output tokens on the international tier.
Token Plan Individual, per QwenCloud’s plan documentation, comes in four tiers paid in Credits:
| Token Plan Individual | Monthly price | New-customer price | Credits per month | Concurrent agents |
|---|---|---|---|---|
| Lite | $6 | $3 | 20,000 | 1-2 |
| Essential | $10 | $5 | 45,000 | 2-3 |
| Standard | $18 | none listed | 80,000 | 3-4 |
| Pro | $43 | none listed | 200,000 | 6-8 |
Extra Credit Packs cost $15 for 20,000 Credits, and you can hold up to five. The new-customer prices end at 23:59:59 Beijing time on November 30 (10:59 a.m. ET that day) and only apply to accounts that have never bought a Token Plan Individual subscription. Alibaba also raised every plan’s monthly quota and cut the Pro price from November 9, the day before the preview opened.
What you can’t do is turn any of that into a per-token figure. QwenCloud says the Credits a request consumes are “dynamically determined by model type, token usage, thinking mode, and tool calls,” and it points subscribers to a usage page in the console rather than publishing a rate card. For the preview, Alibaba says Qwen4-Max calls use “as low as 20% of the standard rate (5x usage)” during a limited-time promotion. With thinking defaulting to xhigh and counting against your output, the only reliable way to learn what a task costs is to run it and watch the meter.
Beijing’s night discount lands on the US workday
There’s a second discount on top. Calls made between 22:00 and 08:00 Beijing time get “an additional 60% off, that is, 8% of the standard rate,” according to the plan documentation. Beijing is 13 hours ahead of New York in November, so that window runs from 9 a.m. to 7 p.m. Eastern, or 6 a.m. to 4 p.m. Pacific. By our arithmetic, 8% of the standard rate stretches a plan’s Credits 12.5 times as far as normal, and 2.5 times as far as Qwen4-Max’s own daytime preview rate.
We’ve seen this before. When we covered DeepSeek V4.1 Flash 1017, its off-peak price also fell on US business hours, so a developer in New York or San Francisco who works a normal day pays the cheap rate. Token Plan gives the same kind of night discount to other models it carries: 60% off Qwen3.8-Max and Qwen3.8-Flash, and 50% off four DeepSeek V4 and V4.1 models, including V4.1 Flash 1017.
The fine print: interactive use only, no cancelling, Singapore servers
The terms are where Token Plan stops looking like a normal API. QwenCloud limits use to “interactive use within programming and agent tools (such as Claude Code, Cursor, Qwen Code, Qoder, OpenClaw, etc.)” and says the plan “must not be used for automation scripts, custom application backends, or any non-interactive batch call scenarios.” Calls outside that scope can get a subscription suspended or an API key banned. In practice, that makes the preview a way to try Qwen4-Max as a coding assistant, not a way to build or test a product on it.
Three more rules are worth reading before you pay. The subscription is for one person, with one per account, and sharing a key can get access restricted. The only region is Singapore with “Global” deployment, which QwenCloud says means inference can run anywhere and prompts and outputs “will involve cross-border data transfer,” with compliance left to the customer. And the Individual plan “does not currently support cancellation.” A month bought on November 11 runs past the preview’s end, though the plan also covers Qwen3.8-Max, GLM-5.2, DeepSeek V4 Pro and other models.
The fixed preview has one genuine upside. Alibaba says “no changes will be made during this period,” so anyone evaluating Qwen4-Max is testing the same snapshot for two weeks. The flip side is that any workflow pointed at qwen4-max-preview stops working when it’s retired at 11 a.m. ET on November 25, the day before Thanksgiving in the US. By our count the window is 350 hours long.
How Qwen4-Max compares on paper with this week’s other 1M-token models
On the specs Alibaba has published, Qwen4-Max isn’t unusual. Every model below reads about a million tokens. What sets it apart is that it’s the only one you can’t price per token and the only one with a shutdown date. Prices come from each company’s own page, checked November 11:
| Model | Price per million tokens (input / output) | Context | Max output | Who can use it |
|---|---|---|---|---|
| Alibaba Qwen4-Max preview | no per-token price (Token Plan Credits) | 1M | 393,216 (thinking + answer) | Token Plan subscribers, November 11-25 Beijing time |
| Google Gemini 4 Carbon | $2 / $10 through November 30, then $4 / $20 | 1,048,576 (shared) | up to 1M minus input | paid Gemini API tier; Google AI Ultra in apps |
| Anthropic Claude Fable 5.2 | $8 / $40 | 1M | 128K | Claude API and paid Claude plans |
| Meta Muse Ember | $0.10 / $0.50 (over 200K input: $0.30 / $1.50) | 1,048,576 | 524,288 | Meta Model API and Muse Code |
Google’s Carbon is the closest parallel, and the contrast is instructive. When Google announced Gemini 4 Carbon on October 28, only vetted cyber defenders in its Fairwind Program could use it, and the end of the $2 / $10 introductory price had no date. On November 10, Google opened Carbon to all paying API developers and Google AI Ultra subscribers and set the date: the price rises to $4 / $20 after November 30. Anthropic’s Claude Fable 5.2, released a day earlier, sits at the expensive end at $8 / $40. Meta’s Muse Ember is the budget option in the same Muse family as Muse Wonder, which costs $4 / $20. All three have a public rate card. Alibaba gave Qwen4-Max an expiry date instead.
We’ve deliberately left benchmarks out of that table. Google and Anthropic published their own; Alibaba published none, and we found no independent results for Qwen4-Max on Wednesday morning. Any “Qwen4-Max vs Gemini” score you see this week comes from someone’s private test, not Alibaba.
What Alibaba hasn’t said about Qwen4-Max
The list of gaps is long for a flagship. Alibaba hasn’t said what Qwen4-Max will cost after November 25, when or whether a non-preview version arrives, or which other models make up “Qwen4.” It hasn’t said how many parameters the model has or what it was trained on. It hasn’t said whether the weights will be released, a question Chinese labs have recently answered late: Moonshot sold Kimi K3.1 Pro as open-weight before the download existed, and Tencent called Hy4 open source while the weights were still “in the coming days.” And the team’s post doesn’t say what the Qwen app, QwenWork, Qoder or QwenStudio charge for early access, or how much use they allow.
What to watch next
Two dates matter. Qwen4-Max-Preview retires at 11 a.m. ET on November 25, and Token Plan’s $3 and $5 new-customer prices end on November 30, the same day Google’s Carbon introductory price runs out. If Alibaba wants developers to commit to Qwen4, it will need to publish a per-token price and some numbers before the preview ends. Until then, nobody can build a product on a model that’s scheduled to switch off the day before Thanksgiving.
Frequently asked questions
How much does Qwen4-Max cost?
Alibaba hasn’t published a per-token price. During the preview, developers can only use it through a Token Plan subscription: Lite at $6 a month, Essential at $10, Standard at $18 or Pro at $43, with Lite and Essential at $3 and $5 for new customers until November 30. Qwen4-Max calls use as little as 20% of the normal credit rate, and 8% between 9 a.m. and 7 p.m. Eastern.
When does the Qwen4-Max preview end?
Alibaba says the preview runs from 10:00 on November 11 to 24:00 on November 25, 2026, Beijing time. That’s 9 p.m. ET on November 10 to 11 a.m. ET on November 25. Alibaba says the model won’t change during the preview and will be retired when it ends.
What is Qwen4-Max’s context window?
It has a 1M-token context window and can produce up to 393,216 tokens per response, but that output budget is shared between the model’s thinking and its final answer. It accepts text, images and video as input.
Can I use Qwen4-Max in my own app?
Not during the preview. Pay-as-you-go access isn’t offered, and Token Plan’s terms limit it to interactive use inside coding and agent tools like Claude Code, Cursor and Qwen Code. They ban automation scripts, custom app backends and batch jobs.
Is Qwen4-Max open source?
Alibaba hasn’t said. Its model page and Token Plan documentation don’t mention downloadable weights or a license, and they give no parameter count.
Sources
- Alibaba Cloud Model Studio: qwen4-max-preview (last updated November 11, 2026; capabilities, 1M context, 393,216 max output, preview dates, Token Plan only, max_tokens change, thinking levels)
- QwenCloud docs: Token Plan Individual overview (plan prices and Credits, new-customer price end date, Qwen4-Max preview Credit rates and night discount, usage scope, Singapore region, cancellation, alias)
- QwenCloud: Token Plan pricing page
- QwenCloud docs: Model pricing (Qwen3.8-Max $2 / $6 international; no Qwen4-Max entry)
- Qwen on X: “Qwen4 is coming,” November 10, 2026, 9:01 p.m. ET
- TechNode: “Alibaba opens Qwen4-Max preview with a million-token context window,” November 11, 2026
- Google: “Gemini 4 Carbon is now available,” November 10, 2026
- Google AI Studio docs: What’s new in Gemini 4 Carbon (pricing, context, long decode)
- Claude Platform docs: Claude Fable 5.2 overview (price, context, max output)
- Meta Model API docs: Muse Ember (context window, max output)
- Meta Model API docs: Pricing and rate limits (Muse Ember tiers)
All sources accessed November 11, 2026.
Last updated: November 11, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
