Skip to content
Chain AI Wire
AI Models

Alibaba Launches Qwen4-Max: 2.4T-Parameter Flagship, 1M-Token Context and Pricing Far Below GPT-6.2

Alibaba's Qwen4-Max pairs a 2.4-trillion-parameter model with a one-million-token context and $1.50 / $4.50 pricing - a fifth of what OpenAI and Anthropic charge for their flagships.

Illustration of Alibaba Qwen4-Max, a glowing AI core above a Hangzhou skyline at dusk with streams of code and a million-token context ribbon
Alibaba's Qwen4-Max pairs a 2.4-trillion-parameter model with a one-million-token context window and prices far below its Western rivals.

At a glance: Alibaba released Qwen4-Max on 16 November 2026, the first flagship of its new Qwen4 family. It carries 2.4 trillion total parameters but activates only 71 billion per token, reads up to one million tokens of context, and costs $1.50 per million input tokens and $4.50 per million output tokens through the QwenCloud API – roughly a fifth of what OpenAI and Anthropic charge for their flagships. Alibaba says the average cost of a finished task falls about 62% against its predecessor, Qwen3.8-Max. That figure comes from the company’s own Work360 benchmark, and independent results have not yet appeared.

Alibaba has released Qwen4-Max, its most capable model to date, and the pitch is as much about price as power. The model went live on QwenCloud on Monday 16 November, after a Token Plan preview that began on 11 November, and it arrives in the middle of a crowded month for frontier models: OpenAI shipped GPT-6.2 Astra the next day, and Google’s Gemini 4 Carbon and Anthropic’s Claude Fable 5.2 landed in the same two-week window. Qwen4-Max undercuts all three on list price by a wide margin, and Alibaba is betting that cost per finished task, not raw capability, is where the next round of the race is won.

A bigger backbone that works cheaper

Qwen4-Max is built on a new sparse mixture-of-experts architecture. The headline numbers, from Alibaba’s launch announcement: 2.4 trillion total parameters, 71 billion activated per token, plus 328 billion N-gram embedding parameters. That is the same 2.4-trillion backbone as Qwen3.8-Max, but with about 25% fewer parameters active on each token, which is where the efficiency gain comes from.

The model takes text, image and video as input and returns text, with native vision built in rather than bolted on. The context window runs to one million tokens, and Alibaba’s Model Studio documentation lists up to 131,072 output tokens in non-thinking mode, or 262,144 tokens shared between thinking and the final answer. For long documents, whole codebases and multi-step agent work, that is the practical difference: the model can hold far more of a job in view at once.

On benchmarks, Alibaba reports a 6th-place ranking on Code Arena, a real-world coding leaderboard, and strong scores on agent and workplace tests including Agents’ Last Exam, Vals Index, JobBench and Chartography. Reuters reported that Alibaba said the model scored 14.7% on a secure-code-generation benchmark, against 8% for its predecessor, while noting those findings had not been independently verified. None of the benchmark figures in the launch have been reproduced by an outside lab yet.

Pricing compared

The price is the story. At $1.50 input and $4.50 output per million tokens in the Singapore region, Qwen4-Max costs less than a fifth of the $8 / $40 charged by GPT-6.2 Astra and Claude Fable 5.2, and undercuts even Gemini 4 Carbon’s $2 / $10 introductory rate. Cached input, billed as an implicit cache hit, costs $0.075 per million tokens. Prices in the Beijing, Hong Kong, Frankfurt, Virginia and Tokyo regions are slightly lower still, at $1.413 input and $4.24 output.

ModelInput / 1M tokensOutput / 1M tokensContext window
Qwen4-Max$1.50$4.501M tokens
GPT-6.2 Astra$8$40Not disclosed
Claude Fable 5.2$8$40Not disclosed
Gemini 4 Carbon$2$101M tokens
Qwen3.8-Max (previous)$2$61M tokens

Alibaba’s framing goes further than list price. On Work360, its own benchmark covering 721 specialist fields, the company says Qwen4-Max uses about 35% fewer tokens at high reasoning effort than Qwen3.8-Max, and that the average cost per task fell from $4.43 to $1.67 – a reduction of around 62%. Work360 is Alibaba’s own evaluation, so the saving should be read as the vendor’s claim until independent testing confirms it. But the direction is consistent with the list-price cut of 25% against the previous flagship, and with the cheaper caching tier.

Built for long-running agents

Alibaba positions Qwen4-Max as a teammate for sustained work rather than single-turn answers, and the launch announcement leans on two showcase runs. In chip cooling research, the model completed a continuous 48-hour agent run through four iterations of a closed-loop workflow – molecular design, synthesis and testing, active learning – to discover an experimentally validated coolant molecule. In an Alzheimer’s research experiment, it executed 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules. Both are company-run demonstrations, not independent results, but they show where Alibaba thinks the model’s edge lies: long-horizon planning that keeps going without a human in the loop.

The same ambition shows up in the product surface. Qwen4-Max is available now through the QwenCloud API and Token Plan, and across Alibaba’s own apps including QwenWork and QwenStudio, as well as the Qoder coding tool. The preview version, which ran on Token Plan only from 11 November, retires on 25 November, leaving the production API model as the only route in. The API is compatible with OpenAI and Anthropic formats, so existing client code can be repointed with minimal changes.

Open weights later, and a closed flagship for now

One thing Qwen4-Max is not: open. Alibaba says smaller models in the Qwen4 family will be released as open-weight models, but it gave no date, and the Max weights stay closed. That stands in sharp contrast to the week’s other big release, K2-Horizon, whose six models from 0.9B to 375B shipped fully open with training data, checkpoints and code. Teams that need to self-host or fine-tune will have to wait for the smaller Qwen4 releases; teams that only need an API have the model today.

Safety: defensive cyber help only

Safety gets unusual prominence in this launch. Alibaba says it filtered training data by task intent, keeping defensive material for vulnerability detection and remediation while removing offensive attack content, and that software-engineering training included code security checks alongside functionality tests. Dedicated safety datasets and post-training work are aimed at chemical, biological, radiological and nuclear risks. Reuters noted that Alibaba’s internal evaluations found no evidence the model could significantly improve a user’s ability to carry out dangerous tasks, and that these findings had not been independently verified. The stance echoes the rest of the industry this month – OpenAI’s GPT-6.2 launch carried a similar critical-level cybersecurity disclosure – but Alibaba is the first to make it a headline feature of the release.

What to watch

  • Independent benchmarks: whether outside labs reproduce the Work360 cost-per-task figures and the Code Arena ranking.
  • The smaller Qwen4 models: when Alibaba releases the promised open-weight models, and under what licence.
  • Real-world agent runs: whether the 48-hour research workflows hold up outside Alibaba’s own demonstrations.
  • Rival pricing: whether OpenAI, Anthropic or Google answer a flagship that costs a fifth of theirs.
  • Preview retirement: the Token Plan preview ends on 25 November, after which only the production API model remains.

What is Qwen4-Max?

Alibaba’s new flagship AI model, launched 16 November 2026. It is a mixture-of-experts model with 2.4 trillion total parameters, 71 billion active per token, native vision and a one-million-token context window.

How much does Qwen4-Max cost?

$1.50 per million input tokens and $4.50 per million output tokens through the QwenCloud API in the Singapore region, with implicit cache hits at $0.075 per million tokens. Prices are slightly lower in the Beijing, Hong Kong, Frankfurt, Virginia and Tokyo regions.

Is Qwen4-Max cheaper than GPT-6.2 Astra and Claude Fable 5.2?

Yes, on list price. Both GPT-6.2 Astra and Claude Fable 5.2 cost $8 input and $40 output per million tokens, roughly five times Qwen4-Max’s $1.50 / $4.50. Alibaba also claims a 62% lower cost per finished task than its own predecessor, though that figure comes from its own Work360 benchmark.

Is Qwen4-Max open source?

No. The Max weights are closed. Alibaba says smaller models in the Qwen4 family will be released as open-weight models, but it has not given a date.

Where can I use Qwen4-Max?

Through the QwenCloud API and Token Plan, and in Alibaba’s QwenWork, QwenStudio and Qoder apps. The API is compatible with OpenAI and Anthropic formats.

David
Written by

David

David is a senior technology and digital assets contributor at ChainAIWire, covering breaking developments across generative AI, autonomous agent ecosystems, and blockchain infrastructure. He tracks how emerging machine learning models intersect with decentralized finance, regulatory policy, and modern market structure. When he isn't analyzing protocol updates and macro liquidity shifts, he closely monitors open-source developer communities.

More articles by this author →

Leave a Comment