Straight answer
● Released 16 November 2026. Closed weights, first model on the Qwen4 architecture - a sparse mixture of experts, 2.4T backbone parameters with 71B active per token.
● The price is the headline $1.50 in / $4.50 out per million tokens. Claude Opus 5.5 is $4 / $20 and GPT-6 Astra is $10 / $50.
● What Alibaba calls it "A long-horizon agent model rather than a chat model" - and their own benchmark chart supports that framing.
● Where it is weak Coding. On Alibaba's OWN chart it loses Terminal-Bench to four rivals and ProgramBench to Opus 5.5 by more than two to one.
Alibaba published a benchmark chart with its launch, and the most useful thing about it is that it shows the model losing. Vendors do not usually do that, and it makes the wins more believable - so this piece works through that chart rather than around it.
Where Qwen4-Max wins on Alibaba's own numbers
| Benchmark | Qwen4-Max | Best rival shown |
|---|
| Agents' Last Exam v1.1 | 46.7 | Opus 5.5 and GPT-6 Astra, both 43.4 |
| HLAB-Vals-1.1 (legal agent work) | 21.3 | Kimi K3.1 11.7; Opus 5.5 9.2 |
| DeepSWE v1.1 | 75.7 | GPT-6.1 Sol 75.2; Opus 5.5 74.2 |
| Work360 (Alibaba's own benchmark) | 52.69 | Qwen3.8-Max 47.74 |
The agentic results are the real ones. Agents' Last Exam and the legal-agent benchmark are both long-horizon tasks, and on HLAB the margin is not close - 21.3 against 9.2 for Opus 5.5 is roughly double. That is consistent with the way Alibaba is positioning the model, and it is the case for trying it.
Two caveats on that table. Work360 is Alibaba's own benchmark, so the only comparison it offers is against its own predecessor. And DeepSWE is a win by half a point over GPT-6.1 Sol, which is noise rather than a lead.
Where it loses, and this is the part that matters for coding
| Benchmark | Qwen4-Max | What beat it |
|---|
| Terminal-Bench 4.0 | 48.5 | Kimi K3.1 67.2, Opus 5.5 59.6, GPT-6 Astra 59.1, GPT-6.1 Sol 56.1 |
| ProgramBench | 31.5 | Opus 5.5 65.0 - more than double |
| Humanity's Last Exam | 51.7 | Opus 5.5 59.8, GPT-6.1 Sol 54.0 |
| Chartography | 54.2 | GPT-6 Astra 71.1 |
| Video-Index | 68.1 | GPT-6 Astra 76.7 |
Some launch coverage has reported that Alibaba claims Qwen4-Max beats rivals on coding. On the chart Alibaba published, it does not. It wins DeepSWE by half a point and loses Terminal-Bench to four different models, one of them by nineteen points. ProgramBench is worse - 31.5 against 65.0 for Opus 5.5.
If you are choosing a model to drive a coding agent, that is the finding. Terminal-Bench measures exactly the kind of multi-step shell-and-filesystem work a coding agent does, and Kimi K3.1 at 67.2 is in a different class on it. Our Codex vs Claude Code comparison covers the tools most people actually reach for here.
Independent numbers, which land lower than the launch chart
Artificial Analysis places it at 43 on its Intelligence Index, ninth overall. Vals ranks it 11th of 38 at 63.20%. Vals also ran Terminal-Bench independently and measured 48.48%, which matches Alibaba's 48.5 almost exactly - so the published figures appear honest, the model is simply mid-table rather than frontier-leading. Testers consistently describe it as slow and verbose.
Pricing, which is where it is genuinely competitive
| Model | Input / 1M | Output / 1M |
|---|
| Qwen4-Max | $1.50 | $4.50 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Kimi K3.1 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Fable 5.2 | $8.00 | $40.00 |
| GPT-6 Astra | $10.00 | $50.00 |
Cached input and explicit cache reads are $0.075 per million. Cache creation is $1.875 per million on the five-minute window or $3.00 on the one-hour. There is a free quota of 1M tokens in the Singapore region for around 90 days, and the Token Plan subscription starts around $6 a month on third-party reports.
Against Opus 5.5 that is 2.7x cheaper on input and 4.4x cheaper on output. For agentic workloads where the model wins the benchmark anyway, that combination is hard to argue with. For coding, you are paying less for a model that measures worse.
Two things in the documentation that do not line up
The output limit is stated two ways. The API documentation lists 131,072 tokens without thinking and 262,144 with it; the model page lists 393K in both modes. Those cannot both be right, and if you are building against a long-output assumption it is worth confirming before you rely on it.
Separately, Alibaba reports that average cost per Work360 task fell from $4.43 to $1.67, about 62%, against Qwen3.8-Max. That drop is larger than the token-count and price reductions alone account for, and the launch material does not explain the gap. It may be real efficiency; it is not shown.
Access and the open-weights question
The API model ID is qwen4-max on QwenCloud, with both OpenAI-compatible and Anthropic-compatible endpoints, which makes swapping it into an existing stack a config change rather than a rewrite. It is also in QwenStudio, QwenWork and the Qoder coding product. Inputs are text, image and video; output is text only. The context window is 1M, with maximum input of 991K tokens, or 983K with thinking enabled, and reasoning able to use up to 262K.
Max is closed-weight and there is no release planned for it. Alibaba says smaller Qwen4 models will be open-sourced later, which is the same pattern as the previous generation.
FAQ
When was Qwen4-Max released?
16 November 2026. A preview, qwen4-max-preview, opened on 11 November to Token Plan subscribers only.
How much does Qwen4-Max cost?
$1.50 per million input tokens and $4.50 per million output on QwenCloud's international list. Cached input and cache reads are $0.075 per million. That is 2.7x cheaper than Claude Opus 5.5 on input and 4.4x on output.
Is Qwen4-Max good at coding?
Less than the launch framing suggests. On Alibaba's own chart it scores 48.5 on Terminal-Bench 4.0 against Kimi K3.1 at 67.2 and Opus 5.5 at 59.6, and 31.5 on ProgramBench against 65.0 for Opus 5.5. It wins DeepSWE v1.1 by half a point, which is within noise.
What is Qwen4-Max actually good at?
Long-horizon agent work. It leads Agents' Last Exam v1.1 at 46.7 and the HLAB-Vals legal-agent benchmark at 21.3 against 9.2 for Opus 5.5. Alibaba describes it as an agent model rather than a chat model.
Are the weights open?
No. Qwen4-Max is closed-weight. Alibaba says smaller Qwen4 models will be open-sourced later.
How big is Qwen4-Max?
A sparse mixture of experts with a 2.4T-parameter backbone, 71B active per token, plus 328B N-gram embedding parameters. Context is 1M tokens.
How does it rank independently?
Artificial Analysis puts it at 43 on its Intelligence Index, ninth overall. Vals ranks it 11th of 38 at 63.20%, and its independent Terminal-Bench run of 48.48% matched Alibaba's published 48.5.