Qwen4-Max Preview: Alibaba's New Flagship for $3, for 14 Days

Alibaba's next flagship model is here, and it comes with a deadline printed on the box. On November 11, 2026, the Qwen team announced Qwen4, its next-generation model family, and opened a preview of the flagship tier, Qwen4-Max. You can try it today for $3, through a subscription called the Token Plan, or free in Qwen Chat. Both doors close at midnight on November 25, when Alibaba says the model will be retired. What you cannot do yet is buy it by the token, read its benchmarks, or download its weights, because none of those exist. This is a two-week test drive, and the honest way to use it is to measure the model on your own work before the window shuts.

What Qwen4-Max Preview actually is

Qwen4-Max Preview is a managed preview, not a launch. The announcement came as a short post from the official Qwen account on X, at 02:01 UTC on November 11, 2026. It runs three lines and not a single number: deeper knowledge, richer world understanding, stronger coding and agentic capabilities, built for real-world tasks from complex coding and knowledge work to front-end builds and professional software workflows. The post names five places to try it: the Token Plan, QwenWork, Qoder, the Qwen app, and QwenStudio.

The model ID is qwen4-max-preview, with an alias, qwen4.0-max-preview, listed in Alibaba's documentation. It is the first of four Qwen4 tiers Alibaba named at its Apsara conference on September 22, 2026: Max, Flash, Plus, and a 27-billion-parameter model meant to run locally. Only Max has an endpoint today. The other three have names and no dates.

That matters because the preview is easy to confuse with the finished model. Qwen4-Max Preview is a time-limited service. Qwen4-Max, the generally available flagship, does not exist yet: it has no model card, no published per-token price, and no pay-as-you-go endpoint. Search results and social posts already mix the two up. Treat every number attached to the preview as a property of a two-week trial, not of the model that follows it.

The 14-day window

Alibaba's documentation states the window plainly: the preview runs from 10:00 on November 11, 2026, to 24:00 on November 25, 2026, China Standard Time. After that, the model is retired. The documentation also states that no changes will be made during the period, so a behaviour you find irritating on day two will still be there on day thirteen.

The window is the same on every access route. Qwen Chat, which opened the same model later on November 11, carries the identical expiry. The chat door feels permanent because chat products usually are. This one is not.

Three outcomes are possible on November 25. The clean one is general availability with a published per-token rate and a pay-as-you-go endpoint, which would make the preview the cheapest due diligence anyone could have done. The second is an extension on the same terms, which would say Alibaba is still capacity-constrained and not ready to meter the model openly. The third is retirement on schedule with silence. Nothing in the announcement favours one over the others, and Alibaba has given no date for the finished model.

What it costs

There is no per-token price for Qwen4-Max Preview, because Alibaba does not sell it by the token. Access runs through the Token Plan, a monthly subscription that buys a fixed allowance of credits. Each model burns credits at a rate Alibaba describes as set dynamically by model type, token usage, thinking mode, and tool calls, and publishes no credits-per-token figure. The only place to learn your real rate is the usage detail in the console.

The individual Token Plan has four tiers. Prices below are the list prices on the official Token Plan page, with the new-user promotion in brackets. The page itself renders its prices in JavaScript, so the figures are taken from two independent reads of the same page, which agree.

At November 11, 2026 market rates, verified at publication (1 USD = 0.749 GBP / 1.421 CAD), the $3 new-user Lite plan is about £2.25 or C$4.26 a month, and the $43 Pro plan is about £32.21 or C$61.10. Top-up bundles cost $15 per 20,000 credits, about £11.24 or C$21.32. Team plans run $20, $75, and $180 per seat per month, about £14.98, £56.18, and £134.82, for 25,000, 100,000, and 250,000 credits, with one user and a dedicated API key per seat. Key sharing is prohibited.

Two discounts apply during the preview. Consumption runs at as low as 20% of the standard credit rate, and calls placed between 22:00 and 08:00 China Standard Time get a further discount that lands at 8% of standard. Two Alibaba pages describe that night step differently, one calling it 40% off and the other 60%, and both arrive at the same 8%, so treat the arithmetic as the fact and the percentage as copy. One outlet reports cuts of up to 90%; that figure does not match the plan page, so check the credit count at checkout rather than trusting either number.

The plan has no free trial. That is stated on the plan page and in the launch coverage. The cheapest way in is the $3 Lite promotion for new users, and the subscription caps what you can spend: there is no meter running past the plan's credits unless you buy a top-up.

Two doors: Token Plan or Qwen Chat

Qwen4-Max Preview reached a second surface on November 11, when it appeared in Qwen Chat, the consumer chat app. Same model, same 1-million-token context, same 393,216-token output ceiling, same retirement at midnight on November 25. The doors differ in what they let you do.

The chat door is the better way to answer one question: does this model write good answers to my prompts? The API door is the only way to answer the question that matters for developers: does it do my job, in my tools, at a cost I can predict? A chat conversation cannot be versioned, so two runs cannot be compared honestly. The API lets you pin parameters and record credit consumption per task.

What the model can do

What Alibaba has published about Qwen4-Max Preview fits in a short list, and every item on it comes from the company's own documentation or announcement, with no independent measurement behind it yet.

Two of those figures need a caution flag. The 1-million-token context and the 393,216-token output ceiling come from Alibaba's documentation, as read by launch-day coverage that captured the docs page. But the Qwen Code client also assigns a 1-million-token allocation to the Qwen4 identifiers in its own configuration, and a client setting is not a server limit. Until Alibaba publishes a model card, treat the context figure as documented, not demonstrated. The same applies to the output ceiling: thinking tokens and answer tokens share the 393,216 budget, so a hard prompt that spends most of its budget thinking can leave nothing for the answer.

The model is positioned for coding and long-running agent work: complex code, knowledge work, front-end builds, and professional software workflows. Those are Alibaba's words. No benchmark, vendor or independent, has been published for Qwen4-Max Preview, and the Artificial Analysis leaderboard carries no Qwen4 entry at all as of November 11. Every "Qwen4 beats X" figure circulating today is either a projection or a different model's number wearing this one's name.

What is missing

The gaps are the story, because they decide what the preview is good for. Five things do not exist yet.

There are no benchmark scores. Alibaba published none, and no independent lab has had time to measure a model that opened this morning. The nearest numbers belong to the predecessor, Qwen3.8-Max, which Alibaba reports at 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench, and 74.8 on CoWorkBench. Those scores set the bar Alibaba has to clear. They say nothing about how good Qwen4-Max is.

There is no parameter count. The nearest architecture statement from the Qwen organisation is Qwen3.8-Flash-Next, published on Hugging Face on August 24, 2026, as an experimental preview of the architecture that will underpin Qwen4. It describes Gated DeltaNet and Qwen Sparse Attention, with roughly 6 billion parameters active per step. It describes a smaller preview model, not this flagship. Anyone quoting a parameter count for Qwen4-Max is quoting something Alibaba has not said.

There are no downloadable weights and no licence. Qwen4-Max Preview is a closed hosted endpoint for fourteen days. The family's smaller models, the 27-billion-parameter tier in particular, are the natural candidates for an open release, because that is where expectations have centred since Alibaba showed the new architecture. Whether the Max stays a closed service is for Alibaba to say. So far it has not.

There is no per-token price and no general-availability date. The preview ends November 25, and the finished model could carry a different price in credits, a per-token rate, or neither. The Token Plan is also fenced in by its terms: interactive use in coding and agent tools only, no automated batch scripts, no embedding in the backend of an in-house application, and no key sharing. An API key does not mean the model can be used like a typical commercial API.

There is also a data question the terms answer bluntly. The Token Plan runs on the Singapore region with global inference, and Alibaba's documentation says prompts and responses may cross borders as part of inference processing. If client documents or personal data sit inside the prompt, that transfer has to be decided before pasting the code. Do not paste confidential code into a preview whose data residency you have not checked.

Qwen4-Max Preview vs the models you can buy today

The preview's real competition is not another preview. It is the models already on a rate card, which you can buy by the token, run in production, and hold to a published price. The table below uses each vendor's list price, as reported in this site's own verified reads.

Read the table as a warning, not a ranking. Five of the six rivals sit at the same $2 input price, and the preview cannot be placed on the same scale because it has no per-token price at all. The $3 Lite plan buys 20,000 credits, not a number of tokens, and Alibaba publishes no conversion between the two. The only honest comparison this week is a head-to-head test on the same task: run Qwen4-Max Preview and Qwen3.8-Max through the same client, the same prompt, and the same tool definitions, then record two numbers, how many credits went and whether the result holds up on the first test run.

Qwen3.8-Max is the fallback worth knowing. It has been generally available since August 3, 2026, at $2.00 per million input tokens and $6.00 per million output on Alibaba Cloud Model Studio's Singapore endpoint, with the same 1-million-token context and text, image, and video input. If the preview retires on schedule with silence, Qwen3.8-Max is what you keep running. Our Kimi K3.1 Pro review and our DeepSeek V4.1 Flash 1017 review cover the other two $2-tier rivals in the same price band, and our Claude Code explainer covers the client most readers will plug the preview into.

How to use the 14 days

Fourteen days is enough time to learn something specific if you decide what you are testing before you start. Three prompts, run twice each, is a better use of the window than a week of open-ended chatting.

First, test the context window on something you already know the answer to. Hand it a long document or a screen recording with a detail planted near the end, and check whether it finds it. A million tokens is a claim about capacity, not about attention, and the only way to tell the difference is a needle you put there yourself.

Second, test the reasoning default against the cheap setting on the same prompt. Reasoning effort takes low, medium, or xhigh, and defaults to xhigh. A model that spends its budget thinking before it answers will produce better hard-prompt output and worse latency on easy ones. If xhigh and medium land in the same place on your task, you have found latency you do not need to pay for. Watch one trap while you do it: starting with this model, max_tokens covers thinking plus the final answer, where every earlier Qwen model applied the cap to the answer alone. A request carried over from Qwen3.8-Max with max_tokens set to 4,096 now buys 4,096 tokens of thinking and whatever is left, which on a hard prompt can be nothing. The failure mode is a truncated response, not an error you can catch.

Third, test video input on something short and specific, such as a UI recording with a known bug in it. Video understanding is the least benchmarked capability on any model card, and the one where a preview is most likely to surprise you in either direction.

Fourth, keep the test inside the terms. The plan allows interactive use in coding and agent tools, not automated batch scripts. Large volumes of automated API calls may violate the Token Plan's terms, so a benchmark harness that fires hundreds of requests is the wrong tool for this preview. Test by hand, write down what you learned with dates and settings, and keep the record: in fourteen days that document is the only asset the preview produces.

Common mistakes to avoid

• Building production on a preview. The model retires on November 25. Anything wired into a backend, a scheduled job, or a customer-facing flow will break, and the terms prohibit that use in the first place.
• Quoting the preview's numbers as the flagship's numbers. There is no model card, no benchmark, and no per-token price. The preview's specs describe a trial, not the finished Qwen4-Max.
• Comparing credit prices to token prices. A $3 subscription and a $2-per-million-token rate are different units. Alibaba publishes no conversion, so any "cheaper than" claim between the two is arithmetic without a rate.
• Trusting the context figure without testing it. A 1-million-token window is documented, not demonstrated. Retrieval quality across a long prompt is exactly what a preview is for measuring.
• Pasting confidential code before checking data handling. Prompts and responses may cross borders during inference. Decide on that before the code goes in, not after.
• Assuming the chat door is permanent. Qwen Chat carries the same November 25 expiry as the Token Plan. The chat tab will not outlive the preview.

Is Qwen4-Max open source?

No, for now it is only a service. Alibaba has not announced downloadable weights or a licence for Qwen4-Max. The family's smaller models, the 27-billion-parameter tier in particular, are the natural candidates for an open release.

How much does Qwen4-Max cost?

There is no per-token price. Access runs through the Token Plan, a monthly credit subscription: Lite $6 list or $3 for new users, Essential $10 or $5, Standard $18, Pro $43, with top-ups at $15 per 20,000 credits. During the preview, consumption runs at as low as 20% of the standard credit rate, and night calls drop to 8%.

When does the Qwen4-Max preview end?

At 24:00 on November 25, 2026, China Standard Time. Alibaba's documentation says the model will be retired after the preview, and that no changes will be made during the period. The same expiry applies to Qwen Chat access.

Can I use Qwen4-Max in Claude Code or Cursor?

Yes, through the Token Plan. The plan supports tools that speak the OpenAI and Anthropic protocols, and Claude Code, Cursor, Cline, OpenCode, Codex, and OpenClaw are among those listed. You need the dedicated base URL and the key from QwenCloud's API keys page, because an existing Model Studio key does not open it.

What happens after November 25?

Three outcomes are possible: general availability with a published per-token rate, an extension on the same terms, or retirement on schedule with silence. Alibaba has given no date for the finished model, so the practical move is to keep Qwen3.8-Max, at $2.00 per million input tokens and $6.00 per million output, as the fallback.

Is Qwen4-Max the same as Qwen4?

No. Qwen4 is the model family Alibaba announced on September 22, 2026, with four tiers: Max, Flash, Plus, and 27B. Qwen4-Max Preview is a two-week trial of the Max tier. Only the Max tier has an endpoint today, and the other three have names and no dates.

Alibaba chose to show Qwen4-Max before showing its numbers, and to charge $3 to anyone who wants to measure it themselves. Until November 25, comparing it with Qwen3.8-Max costs a small subscription and an afternoon of real tasks. After that, the model card is what counts, if it arrives. Try it on work you care about, write down what it did, and decide once the finished model has a price you can actually compare.

Sources

• QwenCloud Token Plan page (official) — Individual and Team tiers, concurrent-agent limits, credit multipliers; prices JS-rendered, figures corroborated by the two launch-day reads below.
• Pasquale Pillitteri, "Qwen4-Max arrives in preview on Alibaba's Token Plan, from $3 a month" (November 11, 2026) — full announcement quote, plan table, 80% discount, Qwen3.8-Max benchmarks.
• XenoSpectrum, "Alibaba's Qwen4-Max Preview Offers a 1M-Token Context Window, but Access Ends Nov. 25" (November 11, 2026) — documentation specs, window dates, team pricing, use restrictions.
• OrcaRouter, "Qwen4-Max Preview in Qwen Chat: a second door, same 14 days" (November 11, 2026) — Qwen Chat access, dedicated base URL, night discount, reasoning and token defaults, docs screenshot.
• OrcaRouter, "Qwen4-Max Preview vs GPT-6.1 Sol: countdown vs rate card" (November 11, 2026) — GPT-6.1 Sol $2/$10 comparison, 20% consumption rate.
• AlphaSignal, "Alibaba Launches Qwen4-Max With Agent-First Coding and 1M Token Support" (November 11, 2026) — sparse-specs framing, Qwen Code 0.26.5 clues, Qwen3.8-Max benchmark table.
• Fello AI, "Qwen 4 Release Date: What We Know So Far" (updated November 2026) — teaser timeline, Apsara roadmap, four tier names.
• CloudZero, "Qwen API pricing in 2026" (updated October 15, 2026) — Qwen3.8-Max $2.00/$6.00 from Alibaba's own Model Studio pricing page.

Exchange rates: November 11, 2026 market rates, verified at publication (1 USD = 0.749 GBP / 1.421 CAD). The @Alibaba_Qwen announcement post of November 11, 2026 is quoted from the Pillitteri read above; no direct post URL is linked because the post page was not opened this run.