Alibaba Launches Qwen4-Max, a 1M-Context Flagship Priced 25% Below Qwen3.8-Max
On this page 6 sections
TL;DR
Qwen4-Max, Alibaba's next-generation flagship, went general availability today, five days into a preview that was scheduled to run through November 25. It takes text, images and video, thinks by default, carries a 1,000,000 token context window, and lists at $1.50 per million input tokens and $4.50 per million output tokens on the international endpoint, 25 percent under Qwen3.8-Max in both directions. The weights stay closed. Independent scoring from Artificial Analysis puts its intelligence mid-pack, so the story you can act on is price and plumbing.
What Alibaba shipped
The model is live on Alibaba Cloud Model Studio as qwen4-max, with deployment scopes in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. Input is text, image and video; output is text. Deep thinking is on by default, with three effort levels (low, high and xhigh) and a default thinking budget of 131,072 tokens. On Model Studio the final answer is capped at 131,072 tokens in non-thinking mode, and thinking plus answer together are capped at 262,144.
The launch cadence is the mildly unusual part. The QwenCloud changelog shows qwen4-max-preview landing November 11 for token-plan subscribers only, with a preview window scheduled to November 25. General availability arrived November 16. Previews slipping late is a genre; one shipping nine days early is not.
Qwen telegraphed the architecture months ago. In August the team released Qwen3.8-Flash-Next, an open-weights model whose card described it as a preview of what Qwen4 would be built on. Today's announcement calls Qwen4-Max a new mixture-of-experts architecture with native vision. Alibaba's documentation pages do not list parameter counts, so neither does this post.
The price cut is the actual news
List prices on the Singapore international endpoint read like a staircase: qwen3.7-max at $2.50 in and $7.50 out, qwen3.8-max at $2.00 and $6.00, qwen4-max at $1.50 and $4.50. That is a 25 percent cut on both sides, generation over generation, and the output price includes thinking tokens, not just the visible answer. The global-scope regions (Virginia, Frankfurt, Tokyo, Hong Kong) list it slightly lower, at $1.413 and $4.24. On Qwen's own Chinese platform the card reads 10 and 30 yuan per million tokens, with cache hits at 0.5 yuan.
The part that compounds is explicit caching. Pay $1.875 per million tokens to write a cache with a five minute life, $3.00 for an hour, and $0.075 for every million tokens read back from it, which is 5 percent of fresh input price. If your agent re-sends the same 40K of system prompt and tool definitions on every call, and it does, this is the difference between a demo and a budget line. Think of it as pinning the manual to the workshop wall: pay a little extra once, then a nickel every time the model glances at it.
Vendor claims versus measurements
Qwen's launch post on X claims Qwen4-Max beats Qwen3.8-Max while using up to about 50 percent fewer tokens, dropping average cost per task by more than 60 percent. Treat both numbers as vendor claims until they survive your own evals. If the token reduction holds, it stacks with the price cut: paying 25 percent less for half the tokens works out to roughly a 62 percent cut per task.
The independent view is sober. Artificial Analysis scores Qwen4-Max at 43 on its Intelligence Index, rank 38 of 81 models, and ranks its cost 44th of 80 with a 95 percent cache discount. Useful, and not the top of anyone's chart. The Hacker News thread pulled roughly 240 points and spent most of them arguing about the gap between the price tag and the index.
Also closed. No fine-tuning, no batch inference, no dedicated deployment on Model Studio, and no weights to download, so the self-hosting crowd can only wave from across the street. If you want open Qwen4-family weights, the August Flash-Next repo is still the artifact.
Why a builder should care
Three practical hooks. First, the API speaks OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope, so harnesses that accept a base URL, including Claude Code and Codex, can point at it without an adapter layer. Second, the cache economics favor exactly the long-system-prompt, many-call pattern that coding agents live in. Third, the region list puts Virginia, Frankfurt and Tokyo endpoints on a Chinese frontier model, which shortens a lot of data-residency reviews.
Chain-of-thought summary and encryption is the sleeper feature. The model returns a summary you can display plus encrypted reasoning you pass back unchanged on later turns, instead of re-feeding raw thinking into the context. On Qwen's own platform the limits run looser still: 393K max output and 5 million tokens per minute, per the model card.
Key Takeaways
- GA November 16, five days after preview started and nine days before the preview window was scheduled to close.
- $1.50/$4.50 per million tokens internationally, 25 percent below Qwen3.8-Max on input and output alike, thinking tokens included in the output price.
- Multimodal input (text, image, video), 1M token context, three reasoning effort levels, explicit caching with reads at $0.075 per million tokens.
- Artificial Analysis Intelligence Index 43, rank 38 of 81; price is the differentiator, and the benchmarks are not.
- Closed weights: no fine-tuning, no batch, no dedicated deployment. Qwen3.8-Flash-Next remains the open Qwen4-family artifact.
- OpenAI and Anthropic API compatible, with endpoints in Singapore, Virginia, Frankfurt, Tokyo, Hong Kong and Beijing.
Sources: Qwen4-Max announcement, Alibaba Cloud Model Studio docs, Model Studio pricing, QwenCloud changelog, qwen4-max-preview doc, Qwen model card, Qwen on X, Artificial Analysis, Hacker News