# Alibaba releases Qwen4 Max at $1.50 per million input tokens

> The 2.4-trillion-parameter model leaves its Token Plan preview for the pay-as-you-go API, with a 1M-token context and no open weights.

URL: https://claudegpt.si/news/qwen4-max-api-price-release  
Published: 2026-11-16T02:56:41.039Z  
Updated: 2026-11-16T06:56:24.105Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Models  
Topics: Qwen, Alibaba, Language models, API pricing  
Publisher: ClaudeGPT Si

## Key takeaways

- Alibaba released Qwen4 Max on November 16, 2026, after a Token Plan-only preview that began on November 11.
- The API costs $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for cached input.
- In Alibaba's own charts it beats Claude Opus 5.5 on long agent tasks but trails it on terminal coding and ProgramBench.

Alibaba's Qwen team released Qwen4 Max on November 16, 2026, the first model of its Qwen4 family, and opened it to everyone on the pay-as-you-go API. [The launch post](https://qwen.ai/blog?id=qwen4-max) describes a mixture-of-experts model, which runs only part of its network for each token: 2.4 trillion parameters in total, 71 billion active per token, with native image and video input and a 1 million token context window.

The price is the headline for developers. On [QwenCloud](https://www.qwencloud.com/models/qwen4-max) it costs $1.50 per million input tokens and $4.50 per million output tokens, below the $2 and $6 that OpenRouter lists for Qwen3.8 Max. Cached input drops to $0.075 per million, from $0.25. That puts a model Alibaba benchmarks against Claude Opus 5.5 and GPT-6 Astra at a fraction of their list prices.

## From a two-week preview to a full release

Alibaba first opened Qwen4 Max on November 11 as a preview called `qwen4-max-preview`. It was available only through Alibaba's Token Plan subscription, with no pay-as-you-go access, and the [preview docs](https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview) say it runs until November 25 and is then retired.

The full model now has its own ID, `qwen4-max`, in [Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/qwen4-max). It is served from Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong, and it speaks the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats. That last one means tools built for Claude can point at it by changing the base URL and model name.

Thinking is on by default at "high" effort. Model Studio lists up to 262,144 tokens for thinking and answer combined, and 131,072 tokens of output with thinking off.

## What Alibaba's benchmarks show

Alibaba compared Qwen4 Max with Qwen3.8 Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 on 16 tests. The scores are Alibaba's own chart, though several come from outside leaderboards (Artificial Analysis, Vals, LM Arena, Mercor), which the footnotes name.

| Benchmark | Qwen4 Max | Qwen3.8 Max | Claude Opus 5.5 | GPT-6 Astra |
| --- | --- | --- | --- | --- |
| DeepSWE v1.1 (agentic coding) | 75.7 | 69.3 | 74.2 | 74.1 |
| Terminal-Bench 4.0 | 48.5 | 38.9 | 59.6 | 59.1 |
| ProgramBench | 31.5 | 28.0 | 65.0 | 55.0 |
| Agents' Last Exam v1.1 | 46.7 | 27.9 | 43.4 | 42.6 |
| Humanity's Last Exam | 51.7 | 43.9 | 59.8 | 53.6 |
| HLAB-Vals-1.1 (legal agent work) | 21.3 | 15.4 | 9.2 | 7.1 |

The picture is mixed, and Alibaba's own chart says so. Qwen4 Max leads on long agent tasks and legal work, and edges out the others on DeepSWE. It trails Opus 5.5 and GPT-6 Astra clearly on terminal coding and on ProgramBench, where it scores less than half of Opus 5.5. Kimi K3.1, an open-weight model, beats it on Terminal-Bench (67.2 to 48.5).

On Alibaba's own Work360 test of industry tasks, Qwen4 Max scored 52.69 against 47.74 for Qwen3.8 Max. Alibaba says it used about 35% fewer tokens at high effort, and that the average cost per Work360 task fell from $4.43 to $1.67.

## How it is built

The launch post describes a new architecture for the whole Qwen4 series. Most layers use Gated DeltaNet, which compresses earlier text into a fixed-size state, and every fourth layer uses a new sparse attention that picks only the relevant parts of the context. Alibaba says this keeps long prompts cheap. It also added 328 billion "N-gram embedding" parameters that look up common short token patterns. Alibaba says the base model beat Qwen3.8 Max's base on every in-house test while activating 25% fewer parameters.

## No open weights for Max

Qwen4 Max is closed, like Qwen3.8 Max before it. The post says smaller Qwen4 models "are coming to open source" but gives no date or sizes.

The price makes this worth testing for anyone running long agent jobs on Claude or GPT. The benchmark gaps on terminal and software tasks say it won't replace them for hard coding yet.

Update, November 16: The first independent score is in. [Artificial Analysis](https://artificialanalysis.ai/models/qwen4-max) gives Qwen4 Max 43 on its Intelligence Index, above the median of 41 for comparable models and 38th of the 81 models it ranks. Running one task of the index costs $3.27 on average at list prices.

## Sources

- [Qwen4-Max: Your New Teammate Has Arrived](https://qwen.ai/blog?id=qwen4-max), Qwen (Alibaba)
- [Qwen4-Max model page](https://www.qwencloud.com/models/qwen4-max), QwenCloud
- [qwen4-max](https://www.alibabacloud.com/help/en/model-studio/qwen4-max), Alibaba Cloud Model Studio
- [qwen4-max-preview](https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview), Alibaba Cloud Model Studio
- [Qwen4 Max on OpenRouter](https://openrouter.ai/qwen/qwen4-max), OpenRouter
- [Qwen4 Max: intelligence, performance and price analysis](https://artificialanalysis.ai/models/qwen4-max), Artificial Analysis
