Seriously, It’s AI Seriously, It’s AIAI News • Tools • Trends Subscribe
Latest
AI News

Alibaba's Qwen4-Max Is Here: Half the Tokens, 62% Cheaper — and One Big Catch

November 16, 2026 By Seriously, It’s AI 0 Comments


How I checked this

I read the Qwen team's launch announcement, the QwenCloud model page for qwen4-max, and the Reuters report on the launch, then compared the new model's pricing, context window and output limits against Qwen3.8-Max on Alibaba's own pages. Every benchmark number below is Alibaba's own — I have not run the model myself, and no independent lab has published a rerun.

What Qwen4-Max actually is

Qwen4-Max is the first model of Alibaba's new Qwen4 family, generally available now after a two-week preview that started on November 11. It is a sparse mixture-of-experts model: a 2.4-trillion-parameter backbone with only 71 billion parameters active per token. In plain terms, a very large model that wakes up a small slice of itself for each token, which keeps running costs down.

It takes text, images and video as input, with native vision built in rather than bolted on. The context window is 1 million tokens, the maximum output is 393,216 tokens, and reasoning is on by default with effort adjustable from low to xhigh. It supports tool calling, structured output and web search, with OpenAI- and Anthropic-compatible API formats. You can try it on the QwenCloud API, Qwen Studio, QwenWork and Qoder.

The price cut is the real story

Alibaba is not really selling a smarter chatbot here. It is selling cheaper work. The API costs $1.50 per million input tokens and $4.50 per million output tokens, with implicitly cached input at $0.075 — 25% below Qwen3.8-Max on input and output, and 70% cheaper on cached input.

The bigger claim is about tokens, not just rates. Alibaba says Qwen4-Max uses up to about 50% fewer tokens than Qwen3.8-Max on the same tasks — roughly 48–49% fewer on Terminal-Bench 4.0, and 24% fewer on its Work360 evaluation at xhigh reasoning effort. Together, that takes Alibaba's average cost per task on Work360 from $4.43 to $1.67, a cut of about 62%. Work360 is Alibaba's own framework for modelling complex professional work. Treat the 62% as Alibaba's claim about its own workload mix — but the direction is clear, and it undercuts the flagship pricing of Claude Sonnet 4.6, which also ships a 1M-token window at higher rates.

Built for long jobs, not quick chats

The launch is framed around a "24/7 AI teammate", and the demos lean into long, multi-step work. In real production settings, the model planned production lines and verified machining work, cutting the manual effort to program one workpiece from about 475 minutes to 24 minutes. In a continuous 48-hour run it screened 2.7 million candidate molecules and directed an automated wet lab to produce a physical sample of a new coolant material for data-centre chips.

These are single showcases run with partner platforms, and the drug candidate still awaits wet-lab validation. They prove a long-horizon agent loop can complete — not that it will complete reliably on your problem.

Why this matters

If you run coding agents or document agents at any scale, token efficiency is the bill. A model that does the same work with half the tokens changes the maths even before the lower per-token price kicks in. It also arrives the same month Tencent's Hy4 launched with vision and computer use, and weeks after Moonshot's Kimi K3.1 Pro pushed open-weight models to 3 trillion parameters. Chinese labs are competing on agent economics now, not just benchmark scores.

The other thing to watch is the open side of this family. Qwen4-Max itself is not open source — the weights stay hosted. But the Qwen team says the open weights of smaller Qwen4 models will be released, and reports suggest Qwen4-27B is expected to run on a 16GB RAM setup. For self-hosters, those smaller models are the ones to watch.

The catches

First, the launch prices are introductory. Digital Trends reports the rates are set to double on December 1, so test now and budget on the higher number until Alibaba confirms otherwise. Second, the headline numbers are in-house: Work360 is Alibaba's own benchmark, and no independent evaluation has been published yet. Third, safety improved but is still modest in absolute terms — on the Last Secure Code Bench, code that is secure at all three granularities rose from 8% to 14.7%. Nearly doubling is real progress, and a reminder that most generated code still is not secure by default.

My verdict: if you run agents, test it this week while launch pricing holds — the token savings are the most concrete part of the announcement. But do not rebuild your stack around it until independent benchmarks land and December pricing is confirmed. Skip it if you need to self-host.

Would you switch your coding agent to Qwen4-Max at these prices? Tell me in the comments.

Sources

Qwen Team, "Qwen4-Max: Your New Teammate Has Arrived", Qwen Blog, November 16, 2026 — https://qwenlm.github.io/blog/
Reuters, "Alibaba unveils new flagship AI model with safety upgrades", November 15, 2026 — https://kwsn.com/2026/11/15/alibaba-unveils-new-flagship-ai-model-with-safety-upgrades/
Qwen (@Alibaba_Qwen), launch announcement on X, November 16, 2026.

Comments

Browse by Category

Stay Ahead with Seriously, It’s AI

Get the latest AI news, trends and expert insights — explained simply, every day.

Follow Us →

Stay updated with AI

One short email a week. No spam, unsubscribe anytime.