Vibeleaderboard
← All Intel
Intel / article

Qwen4-Max now available on AI Gateway

Source
vercel.com
Author
Jerilyn Zheng
Date

Context

Vercel's AI Gateway, a service that lets developers call models from many companies through one API and one bill, added Qwen4-Max, the new flagship model from Qwen, Alibaba's model team. A developer already on the gateway can call it as alibaba/qwen4-max with the same key and code, and Vercel says that after a setup step with its command-line tool, coding agents such as Claude Code, Codex and Cursor can select it. The listing does not add a second host or a lower price. The gateway's endpoint data names one provider, Alibaba, and its public model list gives $1.50 per million input tokens and $4.50 per million output tokens, the prices Qwen set at launch. Vercel says the gateway passes provider prices through with no markup and no platform fee.

The cached-input price is the one that shapes an agent's bill. An agent resends the same long conversation on every turn, and text the provider has already processed can be read from a cache at a lower rate. The gateway lists $0.075 per million tokens for those cache reads, the price Qwen announced for implicit caching, the kind the service applies automatically without the developer setting up a cache, and its endpoint entry marks implicit caching as supported. Writing to the cache is listed at $1.875 per million tokens. When we explained Qwen4-Max on OpenRouter, that listing showed the same $0.075 rate but marked automatic caching unsupported, so of the two listings only Vercel's indicates that Qwen's automatic caching applies.

Some of the listing's labels are Vercel's and do not match what Qwen published. The model list tags Qwen4-Max as "open-weights", but Qwen's launch post promises only for smaller Qwen4 models, not for Qwen4-Max. The listing also gives text and images as the only inputs and caps a single answer at 384,000 tokens, while QwenCloud's own model page, as described in our launch explanation, lists video input as well and a 393K output limit. A developer who needs video or the longest answers should confirm the gateway route supports them before switching. The capability claims in Vercel's post, such as building 3D scenes in Blender, finishing tasks with fewer tokens than earlier Qwen Max models and writing more secure code, come with no test results. Qwen's launch post lists 3D modeling among the tasks the model handles and reports its token savings from its own internal evaluation.

Key takeaways · AI-distilled
  • According to Vercel's changelog, Qwen4-Max can build 3D scenes in Blender and playable games in Unity, Unreal Engine and Godot, running what it builds and using visual feedback to fix camera, animation and lighting issues.
  • For long-running tasks the model ships a usable first version early, then improves it in successive releases so the work can be reviewed and redirected before it finishes. Alibaba is open-sourcing a Live Delivery plugin for this workflow.
  • The changelog says Qwen4-Max uses fewer tokens than prior Qwen Max models for the same tasks and writes more secure code by default, with fewer injection, memory-safety and access-control flaws. It gives no figures for either claim.
  • AI Gateway passes provider pricing through with no markup and no platform fee on inference, including Bring Your Own Key requests. Running npx vercel ai-gateway setup makes the model selectable in Claude Code, Codex and Cursor.
Terms in this piece · Glossary
  • open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
  • reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
More from Jerilyn Zheng
Recommended reads
Comments

Checking sign-in…

Loading comments…