# Alibaba's Qwen4-Max is out: $1.50/$4.50 per million tokens, 1M context, closed weights — what's verified

> Alibaba moved Qwen4-Max from a Token Plan preview to a general API launch on November 16, cutting list prices 25% below Qwen3.8-Max and claiming roughly 62% lower cost per task. The weights stay closed, and every headline benchmark so far is Alibaba's own.

Published Nov 16, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/alibaba-qwen4-max-launch-1m-context-closed-weights-cheaper-tasks-what-is-verified

Alibaba's Qwen team launched **Qwen4-Max** for general API use on November 16, the flagship of its new Qwen4 family. OpenRouter lists the model with a November 16 release date, a **1-million-token context window** and list pricing of **$1.50 per million input tokens and $4.50 per million output tokens**, served only by Alibaba Cloud. It follows a preview that opened on November 11 to subscribers of Alibaba's Token Plan, [per Yotta Labs](https://www.yottalabs.ai/post/qwen-4-max-specs-benchmarks-pricing-how-to-access).

Why it matters: Qwen4-Max is pitched as a cheaper agent and coding model, not just a smarter one. Alibaba's main claim is that it finishes tasks with far fewer tokens than its predecessor, so the cost per completed job falls more than the per-token price does. The trade-off is openness: unlike many earlier Qwen releases, the Max model is hosted-only, with no downloadable weights.

## What you get

| Item | Qwen4-Max (per OpenRouter, AlphaSignal) |
|---|---|
| Model ID | `qwen4-max` |
| Context window | 1M tokens |
| Max output | 393,216 tokens, shared between reasoning and answer |
| Inputs / output | Text, images, video / text |
| Reasoning | On by default; effort low, high or xhigh, or off |
| Tools | Function calling, structured output, web search |
| API compatibility | OpenAI Chat Completions and Responses, Anthropic Messages |
| Where | QwenCloud API, Qwen Studio, QwenWork, Qoder; OpenRouter (Alibaba Cloud only) |
| Weights | Closed — hosted only |

Alibaba hasn't published total or active parameter counts, or details of the new mixture-of-experts design. [Startup Fortune](https://startupfortune.com/alibabas-qwen4-max-beats-qwen38-max-with-half-the-tokens-and-a-62-lower-cost-per-task/) notes that makes the architecture claims impossible to check from outside.

## Price versus Qwen3.8-Max

| Per million tokens | Qwen4-Max | Qwen3.8-Max |
|---|---|---|
| Input | $1.50 | $2.00 |
| Output | $4.50 | $6.00 |
| Implicit cache hits | $0.075 | $0.25 |

Figures as reported by AlphaSignal and Startup Fortune from QwenCloud's price list: a 25% cut on input and output and a 70% cut on cached input.

## Alibaba's performance claims

All of the following come from Alibaba and had not been independently reproduced at the time of writing, as [AlphaSignal](https://alphasignal.ai/news/alibaba-s-qwen4-max-slashes-agent-task-costs-62-with-1m-token-context) points out.

- **Work360** (Alibaba's own benchmark of professional tasks, covering 721 fields): Qwen4-Max scores 52.69, against 47.74 for Qwen3.8-Max, per DEV.
- **Cost per task on Work360:** $1.67 versus $4.43, about 62% lower.
- **Token use:** roughly 24% fewer tokens on Work360 at xhigh effort and 35% fewer at high; about 48–49% fewer on Terminal-Bench 4.0.
- **Coding:** Bloomberg reported that Alibaba says the model beats Claude Fable 5.1 and Kimi K3.1 Pro on coding benchmarks, according to DEV's summary; no scores were given.

One outside signal points the other way. DEV cites the Agent Arena live leaderboard on November 13, where Qwen4-Max at xhigh effort ranked 14th on task success (7.46%), behind Claude Fable 5.2 (first, 28.61%) and Fable 5.1 (sixth, 14.10%). That leaderboard measured the preview, and live arenas test different things from fixed benchmarks, but it's a reason not to take the coding claims at face value. A chart showing an Artificial Analysis score of 53 is circulating, but AlphaSignal says Artificial Analysis hadn't listed the model yet.

## The open-weights question

Alibaba says smaller Qwen4 models will get open weights, with no date. The announced family is Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B, and Yotta Labs describes the 27B as the open-weight tier. For teams that picked Qwen because they could self-host it, the flagship is now an API dependency on Alibaba Cloud. That's a different call on data residency and vendor risk than downloading a checkpoint.

## Who should try it

- **Teams running agents on Qwen3.8-Max:** the switch is a model-ID change at a lower list price. Check that the token savings hold on your own workloads before assuming the 62% figure.
- **Teams paying for Claude or GPT on bulk coding work:** it's worth a bake-off on cost per finished task. Expect lower accuracy on harder agent tasks if the Agent Arena gap holds.
- **Anyone needing self-hosting:** wait for the smaller open Qwen4 models.

What to watch: independent scores from Artificial Analysis and other evaluators, whether the 1M-token window holds up on very long inputs, and when the open Qwen4-27B ships.


## Sources

1. [OpenRouter — Qwen4 Max: API pricing and providers](https://openrouter.ai/qwen/qwen4-max)
2. [AlphaSignal — Alibaba's Qwen4-Max slashes agent task costs 62% with 1M-token context](https://alphasignal.ai/news/alibaba-s-qwen4-max-slashes-agent-task-costs-62-with-1m-token-context)
3. [Startup Fortune — Alibaba's Qwen4-Max beats Qwen3.8-Max with half the tokens and a 62% lower cost per task](https://startupfortune.com/alibabas-qwen4-max-beats-qwen38-max-with-half-the-tokens-and-a-62-lower-cost-per-task/)
4. [DEV Community — Qwen4-Max beat Claude on coding benchmarks, then Alibaba locked the weights](https://dev.to/shivaysinghrajput/qwen4-max-beat-claude-on-coding-benchmarks-then-alibaba-locked-the-weights-l28)
5. [Yotta Labs — Qwen 4 Max Preview: specs, benchmarks, pricing and how to access it](https://www.yottalabs.ai/post/qwen-4-max-specs-benchmarks-pricing-how-to-access)
