Models & labs

Qwen4-Max Leaves Preview: 25% Cheaper than Qwen3.8-Max per Token, 14th in Agents on Arena

On November 16, Alibaba launched the final version of Qwen4-Max, available on a pay-as-you-go basis via API on QwenCloud: $1.50 per million input tokens and $4.50 for output, a 25% drop compared to Qwen3.8-Max. Initial independent benchmarks place it as the top Chinese model on Arena, though behind Chinese rival Kimi K3.1 Pro on Artificial Analysis's index. The weights remain closed.

A sleek geometric balance scale in frosted glass and metal, balancing a glowing sphere with small translucent cubes.
AI-generated illustration

Five days after the preview, the final release is here. Alibaba launched Qwen4-Max this Monday, November 16, the first model in the Qwen4 family, announced on X at 3:01 a.m. Paris time. The major change compared to the November 11 preview: the model is no longer reserved for Token Plan subscribers. It can now be used pay-as-you-go via API, with public pricing, and comes alongside initial independent benchmarks.

What Is Available

According to QwenCloud documentation, the API identifier is qwen4-max. The model accepts text, images, and video as inputs, retains a 1-million-token context window, and reasons by default, offering three levels of reasoning effort: low, high (default), and xhigh. In reasoning mode, thinking and response can reach a total of 262,144 tokens. It is compatible with OpenAI’s Chat Completions and Responses APIs, as well as Anthropic’s Messages API, making it easy to plug into tools like Claude Code, Codex, or OpenCode, which are cited in the documentation.

The model is deployed in Singapore, Beijing, Frankfurt, Hong Kong, Tokyo, and Virginia. However, fine-tuning, dedicated hosting, and batch processing are not supported. Alibaba is also offering it within its Qwen Studio, QwenWork, and Qoder products. The preview model qwen4-max-preview remains a distinct offering with its own conditions.

Prices Drop by a Quarter

This is the most concrete takeaway from the announcement. The model page lists $1.50 per million input tokens, $4.50 for output tokens, and $0.075 for implicit cached input reads. For Qwen3.8-Max, its page lists $2, $6, and $0.25.

Per million tokens Qwen3.8-Max Qwen4-Max Difference
Input $2.00 $1.50 −25%
Output $6.00 $4.50 −25%
Implicit cached input $0.25 $0.075 −70%

For a coding agent that frequently rereads the same repository, caching is what matters most: a query referencing a large, already-cached context costs over three times less than before. For example: a session that reads 1 million tokens (800,000 of which are cached) and writes 200,000 cost 0.40 + 0.20 + 1.20 = $1.80 with Qwen3.8-Max; it now costs 0.30 + 0.06 + 0.90 = $1.26 with Qwen4-Max, a decrease of around 30%, even before factoring in any token efficiency gains.

What Independent Benchmarks Show

Alibaba claims that Qwen4-Max outperforms its predecessor using up to roughly 50% fewer tokens. In its announcement thread, it puts the average cost per task on its Work360 benchmark at $1.67, compared to $4.43 for Qwen3.8-Max. However, Work360 is an internal benchmark created by Alibaba itself, which gives Qwen4-Max a score of 52.69.

Two third-party organizations published their initial numbers on the day of the release:

  • Artificial Analysis gives Qwen4-Max a score of 43 on its Intelligence Index (xhigh reasoning), up from 34 for Qwen3.8-Max. Cost per index task drops to $3.27, down from $10.02 for Qwen3.8-Max: roughly three times cheaper. The model still trails Moonshot’s Kimi K3.1 Pro (46 points, $3.68 per task) and lags significantly behind it in terminal agent capabilities: 48% on Terminal-Bench 4.0, compared to 67% for Kimi K3.1 Pro.
  • Arena ranks it #14 on its Agent Arena leaderboard, based on 14,700 real agent sessions, compared to 39th place for Qwen3.8-Max. It is the top-ranked Chinese model, ahead of Tencent’s Hy4 (#19) and Kimi K3.1 Pro (#21). In web development, it debuts at #6 on Code Arena WebDev with 1,750 points, a preliminary score calculated by Arena’s automated AutoEval tool. These leaderboards reflect user votes and sessions and are expected to shift over the coming days.

Both evaluations align with Alibaba’s claims regarding cost: the efficiency gains are real. They also highlight its limits: Qwen4-Max does not catch up to top US models or to Kimi K3.1 Pro on the Artificial Analysis index.

What Is Missing

Qwen4-Max’s weights have not been released. Alibaba only promises that smaller Qwen4 models will be released as open weights, without providing a timeline or model sizes. The announcement on X also omits the parameter count.

Our take

Alibaba is not trying to take the crown, but rather to deliver the best value among large proprietary models. The 25% price reduction per token, and especially the 70% slash for cache, directly targets coding agents that continuously reread the same context. Independent benchmarks support the trend: roughly three times cheaper per task than Qwen3.8-Max according to Artificial Analysis, alongside a notably higher score. For European teams, the presence of a Frankfurt deployment is also significant. However, Qwen4-Max still trails Kimi K3.1 Pro, its direct Chinese rival, on Artificial Analysis’s index and in terminal agents, and it cannot be fine-tuned. Those waiting for a Qwen4 model they can self-host will have to hold out for the promised smaller open models, which currently have no release date.

Sources

  1. Qwen (X) — Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery Primary source
  2. Qwen (X) — Token efficiency and cost (announcement thread) Primary source
  3. Qwen (X) — Real-world productivity, Work360 (announcement thread) Primary source
  4. QwenCloud — Qwen4-Max (model page and pricing) Primary source
  5. QwenCloud — Latest model: Qwen4-Max (documentation) Primary source
  6. QwenCloud — Qwen3.8-Max (model page and pricing) Primary source
  7. Arena (X) — Qwen4-Max debuts at #14 on Agent Arena Primary source
  8. Arena (X) — Qwen4-Max lands #6 in Code Arena: WebDev Primary source
  9. Artificial Analysis (X) — Alibaba has launched Qwen4 Max Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.