Alibaba's "Qwen4-Max" Arrives — Which Tests Did It Win Against Opus 5.5?
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
China's Alibaba has released a new AI called "Qwen4-Max." You're probably wondering how close it gets to the latest U.S. models and what it costs. This article lays out the known facts with concrete numbers.
According to GIGAZINE's report, Alibaba officially released Qwen4-Max on November 16, 2026. The official blog post is titled "Qwen4-Max: Your New Teammate Has Arrived."
The title frames it as a partner that gets work done alongside you — an AI as a collaborator rather than just a tool.
Notably, a preview version had already launched on November 11. A GIGAZINE article covering the preview describes an example where a 3D block-breaking game was built in about 20 minutes — the result was a single 1,456-line HTML file that was actually playable.
Qwen4-Max uses a MoE (Mixture-of-Experts) architecture with approximately 2.4 trillion parameters. MoE works by combining multiple smaller "expert" AI models, each specializing in different areas.
Parameters are essentially the number of tunable knobs inside an AI's brain. More parameters generally means greater capability, but activating all of them every time makes the system heavy. MoE addresses this by only activating the relevant experts for each task, keeping the model nimble relative to its overall size.
Qwen4-Max comes with built-in image recognition and supports a context window of 1 million tokens — the amount of text it can process in a single session. Inputs can be text, images, or video; output is text.
A context this large opens up use cases like feeding in an entire large codebase at once.
Alibaba claims Qwen4-Max uses up to approximately 50% fewer tokens compared to the previous Qwen3.8-Max, meaning lower average costs per task. This is based on the developer's own announcement; no independent verification has been found at this time.
Alibaba released comparison results against Qwen3-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra, and Kimi K3.1.
In this comparison, Qwen4-Max outperformed Claude Opus 5.5 and GPT-6 Astra on multiple tests — for example, DeepSWE v1.1, a benchmark measuring AI coding ability. However, these figures were self-reported by Alibaba.
Let's also look at third-party metrics.
In short, Qwen4-Max does not outperform U.S. models across the board. A more accurate picture is that it has entered the upper tier in coding and agentic tasks — AI that plans and executes multi-step workflows on its own.
API pricing per 1 million tokens is as follows:
Competitor pricing was not confirmed in this research, so it cannot be stated definitively that Qwen4-Max is cheaper than alternatives.
The preview version was available through Alibaba's subscription API "Token Plan," with new users able to get started from around $3. AlphaSignal reports that it can also be connected to OpenAI-compatible and Anthropic-compatible tools such as Claude Code, Cursor, and Cline.
The preview period runs through November 25. Please refer to official documentation for details on how the preview relates to the full release.
Alibaba released demonstrations of Qwen4-Max actually operating tools.
The first shows it controlling Blender, a 3D tool, to create 3D models. The second shows it working with game engines such as Unity, Unreal Engine, and Godot to develop games.
Consider, for example, a solo game developer. Handling 3D character modeling, stage programming, and everything else alone is a heavy lift. An AI that can physically carry out tasks could significantly speed up prototyping.
It also seems well-suited to the same kind of use case as the preview's block-breaking game — producing something functional in a short amount of time.
The development team has announced plans to release Qwen4 series models as open weights — meaning the model internals will be made public and runnable on personal PCs or private servers. Smaller models are expected to follow at a later date.
No specific schedule or model sizes have been announced. Some coverage during the preview phase speculated that a smaller version called Qwen4-27B might be a candidate for release, but this remains speculation.
For readers in Japan, two points are particularly relevant:
Performance in Japanese and availability conditions for users in Japan (such as a Tokyo region) were not confirmed in this research. Please check official documentation before adopting the service. Also note that using this service involves sending data to a Chinese company, so verifying compliance with your organization's policies is essential.
It wins on certain benchmarks, but cannot be called comprehensively superior. Its Artificial Analysis index score is 43, which falls short of models like Claude Fable 5.2.
Reports indicate the preview version was accessible under a free plan through the Qwen chat interface. Free-tier availability for the full release has not been confirmed.
A later release is planned, but no date has been announced. Until an announcement is made, treat it as a tentative plan.
Per 1 million tokens: $1.50 for input, $4.50 for output, and $0.075 for cached input.
Start by checking the official pricing and available regions to see whether it fits your workflow.
This article is a cross-post from AI Friends.