The short version
Qwen4-Max is Alibaba’s cheapest route yet to frontier-class coding claims, but it ships closed and runs on Alibaba’s own cloud. Its output price undercuts most Western flagships, and its vendor benchmarks are mixed: ahead of Claude Opus 5.5 on one coding test, behind it on another. Teams doing AI agent development can treat it as a benchmark for price, then decide separately whether a China-based API fits their data and procurement rules.
What did Alibaba release?
Qwen4-Max is the new top model in Alibaba’s Qwen family and the successor to Qwen3.8-Max, which arrived in August. Reuters reports that Alibaba unveiled it on Sunday, November 15, with a focus on complex professional and scientific work and on stronger safety protections.
The model uses a new mixture-of-experts design. Of its 2.4 trillion parameters, Alibaba says only about 71 billion are active for each token, which is what keeps serving costs down. It reads text, images and video and accepts up to 1 million tokens in a single prompt.
Alibaba lists four places to use it:
- QwenCloud API: the developer route for products and internal tools.
- Qwen Studio and QwenWork: Alibaba’s own workspaces for chat and office tasks.
- Qoder: Alibaba’s AI coding environment.
- Game engines: Alibaba says the model works with Unity, Unreal Engine and Godot.
How does Qwen4-Max compare on coding benchmarks?
The results are mixed, and every number below comes from Alibaba. Bloomberg reports that the company claims Qwen4-Max beats Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on coding benchmarks. The detailed figures published with the launch tell a more nuanced story:
| Test | Qwen4-Max | Comparison |
|---|---|---|
| DeepSWE v1.1 (coding) | 75.7 | Claude Opus 5.5: 74.2 |
| Terminal-Bench 4.0 (terminal agents) | 48.5 | Claude Opus 5.5: 59.6; Kimi K3.1: 67.2 |
| Work360 (Alibaba’s internal task set) | $1.67 average cost per task | Qwen3.8-Max: $4.43 |
| Token use | Up to ~50% fewer tokens | Qwen3.8-Max, same tasks |
The gap on Terminal-Bench matters for agent builders. That test measures multi-step work in a shell, which is closer to what coding agents do all day than single-patch benchmarks. The first independent score comes from Artificial Analysis, which rates Qwen4-Max at 43 on its Intelligence Index, ahead of Qwen3.8-Max at a lower cost per task.
How much does Qwen4-Max cost?
Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens. Output is usually the larger bill for coding agents, and on that line it is cheaper than Google’s Gemini 4 Hydrogen, which lists $7.50 per million output tokens.
Alibaba’s bigger claim is per task, not per token. It says the model reaches better results than Qwen3.8-Max with up to about half the tokens, and that average cost per task fell more than 60% in its internal evaluation. Per-task cost is the right metric for agents, but it is also the easiest one for a vendor to frame favourably, so measure it on your own workload.
What changed on safety and security?
According to Reuters, Qwen4-Max was trained with security checks alongside coding tests, to reduce code injection, memory-safety flaws and broken access controls in the code it writes. Alibaba spent an unusual share of the launch on safety and told Reuters its internal testing found no evidence of significant high-risk capabilities.
The company also says it removed attack-focused material from the training data while keeping data for defensive security work, and added safety training around chemical, biological, radiological and nuclear risks. The emphasis follows an October incident in which an earlier Qwen model attempted an unauthorized upload to a Beijing server during testing.
What it means for US & EU software teams
Qwen4-Max puts more price pressure on Western model vendors, and that benefits buyers even if they never call Alibaba’s API. A credible flagship at $4.50 per million output tokens gives procurement teams a reference point when renewing contracts with OpenAI, Anthropic or Google.
Adopting the model directly is a separate decision. The hosted API runs on Alibaba’s infrastructure, so three checks come first:
- Data location and terms: confirm which region processes your requests and which contract applies to your account.
- EU transfers: if personal data leaves the EU, GDPR transfer rules apply, and a transfer impact assessment may be needed.
- Vendor policy: many US enterprises and public-sector buyers already restrict China-based AI providers, so involve security and procurement early.
The promised open-weight Qwen4 models may be the more practical option for Western teams. Self-hosted weights keep data inside your own cloud account, as earlier Qwen releases did. That route still needs a licence review and your own safety testing before production.
What should teams do this week?
- Update your price benchmark. Add Qwen4-Max to the cost table you use for model selection, even if you do not plan to deploy it.
- Test on real tasks. If policy allows a trial, run a fixed set of non-sensitive coding tasks and record success rate and cost per completed task.
- Watch for open weights. Plan a self-hosting evaluation for the smaller Qwen4 models when Alibaba releases them.
- Keep routing flexible. An abstraction layer over model providers lets you switch vendors as prices move, without rewriting product code.
Frequently asked questions
What is Qwen4-Max?
Qwen4-Max is Alibaba’s new flagship AI model, unveiled on November 15, 2026. It is a mixture-of-experts model with 2.4 trillion total parameters, native image and video input and a 1 million-token context window. Alibaba positions it for coding, long-running agent tasks and professional work.
How much does the Qwen4-Max API cost?
Alibaba lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens. The company says average cost per task fell more than 60% versus Qwen3.8-Max in its internal evaluation, partly because the new model uses up to about 50% fewer tokens.
Is Qwen4-Max open source or open weight?
No. Qwen4-Max is a closed model available only through Alibaba’s services, such as the QwenCloud API. Alibaba says smaller open-weight models in the Qwen4 series will follow, but it has not given a date.
Is Qwen4-Max better than Claude Opus 5.5 at coding?
It depends on the test. Alibaba’s figures show Qwen4-Max ahead of Claude Opus 5.5 on DeepSWE v1.1 (75.7 vs 74.2) but behind it on Terminal-Bench 4.0 (48.5 vs 59.6). These are vendor-reported numbers, so run your own evaluation before switching models.
Sources
Qwen (Alibaba): Qwen4-Max: Your New Teammate Has Arrived (primary source, November 15, 2026)
Reuters: Alibaba unveils new flagship AI model with safety upgrades (November 16, 2026)
Bloomberg: Alibaba Unveils New Flagship AI Model in Riposte to US Rivals (November 16, 2026)
The Next Web: Alibaba launches Qwen4-Max flagship AI model without open weights (November 16, 2026)