Qwen 4 Max costs $1.50 per 1M input tokens and $4.50 per 1M output tokens on Alibaba Cloud's international price list. Cached input is $0.075. The model went on sale on November 16, 2026 under the model ID qwen4-max, and it is priced 25% below Qwen3.8 Max, the flagship it replaces.
That is the part worth a buyer's attention. Alibaba did not hold its flagship rate at $2.00 and add a new tier on top. It released a stronger model and set the rate lower. This page covers the official rates, what a real workload pays, who sells the model, and where it sits against the other flagship models on the market.
Qwen 4 Max official API pricing
Alibaba publishes two sets of rates. The international list is served from Singapore. A second, lower set applies on the Beijing endpoint and under what Alibaba calls the Global deployment scope, which covers its Hong Kong, Frankfurt, Virginia and Tokyo regions.
| Rate (USD per 1M tokens) | International (Singapore) | Global scope and Beijing |
|---|---|---|
| Input | $1.50 | $1.413 |
| Cache read | $0.075 | $0.071 |
| Cache write, 5 minutes | $1.875 | $1.767 |
| Cache write, 1 hour | $3.00 | $2.827 |
| Output, thinking included | $4.50 | $4.24 |
Thinking and non-thinking modes share one price. The context window is 1M tokens, with a maximum input of 991,808 tokens and a maximum output of 131,072 tokens, or 262,144 when thinking is on. The model accepts text, image and video input.
The international list includes a free quota of 1 million tokens, valid for 90 days. The two deployment scopes differ in where inference can run, so check Alibaba's deployment scope documentation before choosing one on price alone. The rest of this page uses the international rate.
Qwen 4 Max vs Qwen3.8 Max: a new flagship at a lower rate
Qwen3.8 Max went generally available on August 3 at $2.00 input and $6.00 output. Qwen 4 Max arrives fifteen weeks later at $1.50 and $4.50. Input and output are both 25% lower.
The cached rate moved more. Qwen3.8 Max charges $0.25 per 1M cached input tokens, which is 12.5% of its input price. Qwen 4 Max charges $0.075, which is 5% of its input price. For an agent or a retrieval workload that resends the same long prefix on every call, that line matters more than the headline rate.
It also puts Alibaba's flagship back near where it was before. Our Qwen API pricing page recorded Qwen3.7 Max at $1.475 input and $4.425 output in September. The $2.00 rate on Qwen3.8 Max lasted one model generation.
What changed in between is the competition. Mistral Large 4 reset to $0.70 input in October. Kimi K3.1 Pro arrived in October at $2.00 input and has since opened its weights. Tencent Hy4 launched on November 4 at a promotional $0.417. Better models create competition, and a flagship rate set in August had to answer to the models released in October and November.
What a real workload pays
Take a production workload of 10M input tokens and 1M output tokens per day on the international list.
| Scenario | Qwen 4 Max per day | Qwen 4 Max per year | Qwen3.8 Max per day | Qwen3.8 Max per year |
|---|---|---|---|---|
| No caching | $19.50 | $7,118 | $26.00 | $9,490 |
| 60% of input read from cache | $10.95 | $3,997 | $15.50 | $5,658 |
Without caching, the same workload costs $6.50 less per day on Qwen 4 Max, or $2,373 less per year. With 60% of input read from cache, the gap is 29% instead of 25%, because the cache read rate dropped further than the input rate. The cached figures leave out cache write charges, which are billed at 125% of the input price for a 5-minute cache and 200% for a 1-hour cache.
Under the Global deployment scope, the uncached workload is $18.37 per day, or $6,705 per year.
Thinking, caching and batch: what moves the bill
A few details change what a team actually pays.
Thinking tokens are billed as output. Alibaba's output price covers the chain of thought plus the answer, and the default reasoning effort is high. If thinking adds three tokens for every answer token, the 1M output line in the workload above becomes 4M and the daily bill goes from $19.50 to $33.00. Set reasoning effort per task instead of leaving the default on everything.
Caching needs a prefix of at least 1,024 tokens. Alibaba supports both automatic and explicit caching on Qwen 4 Max, with a 5-minute or 1-hour lifetime. The 1-hour write costs twice the input price, so it only pays off when the same prefix is read many times within the hour.
Batch inference is not supported at launch. Qwen3.8 Max does have a batch route, and OpenRouter lists it at $1.15 input and $3.45 output. For offline work where latency does not matter, the older model on batch is still the lower rate.
Rate limits scale with spend. Alibaba lists 5M tokens per minute for accounts spending up to $8,000 a month, 10M for accounts between $8,000 and $140,000, and 20M above that.
Who sells Qwen 4 Max
One company. Alibaba has not released Qwen 4 Max weights, and we found no Qwen 4 repository under the Qwen organization on Hugging Face as of November 16. OpenRouter lists a single endpoint, served by Alibaba.
That endpoint is priced at $1.425 input and $4.275 output, with cache reads at $0.0713. That is exactly 5% below Alibaba's international list, and it is not flagged as a promotion. It sits between Alibaba's two official rates. Budget against the official list until the reason for the gap is clear.
The preview is going away. Qwen4-Max-Preview was available through Alibaba's Token Plan only, with a preview period running from November 11 to November 25, 2026, Beijing time. Alibaba says it will be retired when the preview ends. Anything built on the preview model ID needs to move to qwen4-max before then.
If weights do arrive, Qwen's own history says not to expect a wide spread. When Alibaba opened the Qwen3.8 2.4T A95B weights in August, six of the seven sellers we tracked, Alibaba included, listed it at exactly $2.00. Why token prices differ covers when open weights produce a spread and when they do not.
Qwen 4 Max against the market
The same 10M input and 1M output workload at each developer's official rate, uncached.
| Model | Input | Output | Per day | Per year |
|---|---|---|---|---|
| Tencent Hy4 (promo through Dec 31) | $0.417 | $1.2505 | $5.42 | $1,978 |
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | $8.58 | $3,132 |
| Mistral Large 4 | $0.70 | $2.10 | $9.10 | $3,322 |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | $17.16 | $6,263 |
| GLM-5.2 | $1.40 | $4.40 | $18.40 | $6,716 |
| Qwen 4 Max | $1.50 | $4.50 | $19.50 | $7,118 |
| Qwen3.8 Max | $2.00 | $6.00 | $26.00 | $9,490 |
| Kimi K3.1 Pro | $2.00 | $10.00 | $30.00 | $10,950 |
Qwen 4 Max lands in the middle. It is priced a little above GLM-5.2 and DeepSeek V4 Pro at peak hours, a little more than twice Mistral Large 4, and well below Kimi K3.1 Pro on output. Unlike DeepSeek, Alibaba charges one rate at every hour of the day. Unlike Tencent, the rate is not a promotion with an end date.
The LLM API pricing comparison has the full set of models side by side.
What it means for buyers
If you run Qwen3.8 Max today, the context window and the 131,072 token output limit are the same on Qwen 4 Max, so the move is a model ID change and a regression test. At 10M input and 1M output per day, it is worth $2,373 a year before caching. If you pay for Qwen3.8 Max Prime at $4.00 and $12.00, check whether you still need the speed tier, because OpenRouter lists no Prime version of Qwen 4 Max yet.
If you are choosing a flagship model for the first time, Qwen 4 Max has one seller and no second source. Kimi K3.1 Pro and DeepSeek V4 Pro both have open weights and many hosts. That matters for uptime, for negotiating and for what happens if the rate changes.
And the rate does change. Across three flagship models this year, Alibaba's input rate has gone from about $1.48 to $2.00 to $1.50, each time on the vendor's schedule. A rate card is one seller's current offer. The Mercatus Token Index tracks where these models actually trade, and a forward contract on the Mercatus Forward Market reserves a month of inference capacity at a price set by the market instead of by one vendor.
Frequently asked questions
How much does the Qwen 4 Max API cost?
$1.50 per 1M input tokens, $0.075 per 1M cached input tokens and $4.50 per 1M output tokens on Alibaba Cloud's international list. The Beijing endpoint and the Global deployment scope are $1.413 input and $4.24 output.
How does Qwen 4 Max pricing compare to Qwen3.8 Max?
Qwen 4 Max is 25% lower on input and output: $1.50 and $4.50 against $2.00 and $6.00. The cached input rate is 70% lower, $0.075 against $0.25.
Is Qwen 4 Max open weight?
No. As of November 16, 2026, Alibaba has not released the weights, and Alibaba is the only seller.
Are thinking tokens billed on Qwen 4 Max?
Yes. Thinking tokens are billed at the output rate of $4.50 per 1M tokens, and thinking and non-thinking modes share the same price.
Why does OpenRouter show $1.425 for Qwen 4 Max?
OpenRouter's single Alibaba endpoint is listed at $1.425 input and $4.275 output, 5% below Alibaba's international list. The listing does not give a reason and is not flagged as a discount.
What happens to Qwen4-Max-Preview?
Alibaba's documentation says the preview period ends at 24:00 on November 25, 2026, Beijing time, and the preview model will be retired after that. The preview was available through Token Plan only.
Does Qwen 4 Max have a free tier?
Alibaba's international list includes a free quota of 1 million tokens, valid for 90 days.
Methodology
Official prices were read from Alibaba Cloud Model Studio's model pricing, Qwen 4 Max model page and explicit cache documentation on November 16, 2026. Third-party prices were read from OpenRouter's endpoint data the same day. Qwen3.8 Max rates are Alibaba's international list. Other models use the official rates on their Mercatus pricing pages. Workload figures assume 10M input and 1M output tokens per day for 365 days. Last verified: 2026-11-16.
