Alibaba Releases Qwen4-Max, Pricing It at $1.50 per Million Input Tokens in Singapore

Alibaba on Monday, Nov. 16, 2026, released Qwen4-Max, a 2.4-trillion-parameter model with a Singapore list price of $1.50 per million input tokens via API.

Share this article

Featured image: Alibaba’s headquarters in Hangzhou, China, in an April 2012 file photo. Photo: Thomas Lombard, Wikimedia Commons, CC BY-SA 3.0. Cropped to 16:9.

Alibaba on Monday, Nov. 16, 2026, released Qwen4-Max, a new flagship model it said has 2.4 trillion total parameters, native vision and a one-million-token context window. In a Hangzhou announcement posted on Alibaba Cloud’s blog, the company called it “the most capable model in the Qwen family to date.”

Advertisement

On QwenCloud’s public pricing page, the Singapore list price for qwen4-max is $1.50 per million input tokens and $4.50 per million output tokens. That input price matches the U.S. list price Google published last week for Gemini 4 Hydrogen; the Qwen4-Max output price is lower. Those are list rates, and they vary by region.

2.4 trillion / 71 billion
Total parameters, and parameters Alibaba said the mixture-of-experts model activates per token
$1.50 / $4.50
Singapore list price, per million input and output tokens, on QwenCloud’s pricing page
25% lower
Boston Report calculation comparing that Singapore list price with Qwen3.8-Max’s $2 and $6 rates
Nov. 16, 2026
Date Alibaba published the Hangzhou announcement

Parameter counts and the release date are from Alibaba’s Nov. 16 blog post. Prices are from QwenCloud’s “Model pricing” page, Singapore tab, retrieved Nov. 16, 2026. The 25% comparison is a Boston Report calculation from those list prices.

What Alibaba said Qwen4-Max can do

Alibaba said the model is a sparse mixture-of-experts system that “activates only 71 billion per token — 25% fewer activated parameters compared to Qwen3.8-Max at the same 2.4T scale.” It said the new Qwen4 architecture changes attention, residual connections and embeddings, and that the result is “stronger capabilities at a lower inference cost.” Those performance claims are Alibaba’s. The Boston Report has not independently benchmarked the model.

The company said Qwen4-Max “ranked 6th on Code Arena, a real-world coding leaderboard,” and that it posted strong scores on Agents’ Last Exam, the Vals Index, JobBench and Chartography. It did not publish those numeric scores in the blog post. On Work360, which Alibaba described as its own benchmark of “challenging industry tasks across 721 specialized fields,” the company said Qwen4-Max outperforms Qwen3.8-Max overall while using 35% fewer tokens at a “High effort” reasoning mode and cutting the average cost per task by about 60%. Work360 is a proprietary test, so those percentages cannot be checked against a public leaderboard from the announcement alone.

Alibaba said the model is available now on QwenCloud through a Token Plan and an API, and on company platforms including QwenWork and QwenStudio. It also said “open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide,” without giving a date. The blog’s only price in the announcement text was an implicit-cache-hit input rate of $0.075 per million tokens, which matches the cache-read column on QwenCloud’s pricing table.

The release is the latest in a line of large Chinese models. The Boston Report previously covered Tencent’s Hy4 open-weight model and an earlier report that Apple had brought Qwen models to China. Monday’s model is a new, hosted flagship, not an open-weight release.

By the numbers: list prices, and how they compare

QwenCloud’s pricing table, retrieved Monday, lists qwen4-max in Singapore at $1.50 per million input tokens and $4.50 per million output tokens, for both non-thinking and thinking modes, with a 1-million-token request cap in that tier. The same table lists qwen3.8-max at $2 and $6. By The Boston Report’s calculation, the new Singapore list rates are 25% lower on both input and output (1.50 divided by 2, and 4.50 divided by 6). Cache reads are listed at $0.075 per million tokens in Singapore, and cache writes at $1.875 for a five-minute cache and $3 for a one-hour cache.

In other regions on the same page, including a “Global” deployment tab, qwen4-max is listed at $1.413 per million input tokens and $4.24 per million output tokens. Customers are billed at the rate for the region whose endpoint they call, the documentation says, so a single headline price does not apply worldwide.

Bar chart of official list prices per million tokens: Qwen4-Max in Singapore at $1.50 input and $4.50 output, Qwen3.8-Max in Singapore at $2 and $6, and Google’s Gemini 4 Hydrogen at $1.50 and $7.50
Sources: QwenCloud, “Model pricing,” Singapore tab for qwen4-max and qwen3.8-max, retrieved Nov. 16, 2026; Google, “Introducing Gemini 4 Hydrogen,” Nov. 12, 2026, updated Nov. 13, 2026. List prices only; regions differ. Chart: The Boston Report.

Those Singapore input rates are the same as the Gemini 4 Hydrogen API price Google published on Nov. 12, $1.50 per million input tokens. Hydrogen’s published output rate is $7.50 per million tokens. A Boston Report calculation from the two official output prices is that the Qwen4-Max Singapore list output rate is 40% lower than Hydrogen’s (4.50 divided by 7.50 equals 0.60). That comparison is of published list prices only. It is not a finding about quality, speed, discounts, or what any given customer will pay, and the two models are sold by different companies in different regions.

Advertisement

Efficiency claims Alibaba published, and their limits

Alibaba said Qwen4-Max can optimize its own cloud scheduling and, in benchmark testing, reduced “data-annotation model call costs by an average of 9%” and accelerated “offline CPU rescheduling speeds by more than 14 times.” In an automotive manufacturing showcase, it said the model cut “manual workpiece programming time by 95% — from roughly 475 minutes to just 24 minutes.” A Boston Report check of those two times is 24 divided by 475, or about 5.05% of the original time, a reduction of about 94.95%, which matches the company’s rounding to 95%. The figures come from a company showcase, not from a named factory’s audited results.

The blog also described long-running scientific agents. Alibaba said one 48-hour run produced “an experimentally validated coolant molecule,” and that in an Alzheimer’s research experiment the model “autonomously executed 454 tool calls and 3,239 metric evaluations to design and optimize 24 candidate molecules.” Those episodes are Alibaba’s accounts. The post does not identify a peer-reviewed paper or an outside laboratory confirmation for either result.

On safety, Alibaba said it included code-security checks in training, filtered out “offensive attack content” from cybersecurity data while keeping defensive material, and used dedicated datasets against chemical, biological, radiological and nuclear risks. It did not publish numeric safety scores in the announcement.

Monday’s release adds another large, lower-priced hosted model to a crowded market. Readers can compare recent systems in our AI models guide and follow continuing coverage in our AI news center. Mint and The Peninsula Qatar also reported the launch on Nov. 16, relying on Alibaba’s announcement.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up