Hot AI News

Qwen4-Max cuts API prices 25% and claims 60% lower cost per task

Alibaba's Qwen4-Max uses a new mixture-of-experts design with native vision and a 1M-token window, and its API undercuts Qwen3.8-Max on every price line.

Qwen4-Max cuts API prices 25% and claims 60% lower cost per task
Image: Alibaba (Qwen)

Alibaba’s Qwen team has released Qwen4-Max, a flagship model built on a new mixture-of-experts architecture with native vision and a 1M-token context window. In a post on X, the team said the API is live and priced below Qwen3.8-Max.

Pricing, access and what Alibaba says it can do

The listed prices are $1.5 per million input tokens, $4.5 per million output tokens and $0.075 per million tokens for implicit caching. Qwen says all three are lower than Qwen3.8-Max, and that the caching rate is down by 70%.

Qwen4-Max cuts API prices 25% and claims 60% lower cost per task
Image: Alibaba

The API is available on QwenCloud. Alibaba also lists Qwen Studio, QwenWork and Qoder as places to try the model. The post links a blog, a model page and a Token Plan pricing page on qwen.ai and qwencloud.com.

On quality, the team claims Qwen4-Max beats Qwen3.8-Max while using up to roughly 50% fewer tokens, and that the average cost per task falls by more than 60%. Alibaba attributes both figures to its internal evaluation. The post does not name the tasks, the benchmarks or the settings behind them.

Alibaba describes the gains as better handling of professional tools, stronger long-horizon execution and higher-quality deliverables. It lists 3D modeling, game development, manufacturing, finance, law, healthcare and frontier scientific discovery as example workloads. No scores are given for any of them in the post.

The team also said that open weights of smaller Qwen4 models are planned, without a date or sizes. The model itself is not open. Five days earlier, on November 11, the team had teased that Qwen4 was coming, with deeper knowledge and stronger coding and agentic abilities.

Advertisements

Numbers in context: A 25% price cut and a bigger drop per task

Third-party price trackers put Qwen3.8-Max at $2.00 per million input tokens and $6.00 per million output tokens on Alibaba’s own API. One of them, CometAPI, lists cached input at about $0.25 per million. Those figures come from trackers rather than from the Qwen4 announcement, and the rates were reported as of August 2026.

Measured against them, Qwen4-Max’s list prices are 25% lower for both input and output. The caching rate falls from about $0.25 to $0.075, which matches the 70% reduction Alibaba states. That is a useful cross-check on the announcement, though it only confirms the arithmetic and says nothing about quality.

The per-task claim is a separate matter. A model that needs up to half the tokens at a price 25% lower would cost roughly 62% less per task at best. That is close to the “over 60%” Alibaba gives, so the two claims are consistent. Real savings will depend on how many tokens a given workload uses. Agentic jobs that rely heavily on cached context may gain more from the caching rate than from the headline prices.

The 1M-token window is not new for the family. The same trackers list Qwen3.8-Max with a 1,000,000-token window and up to 131,072 output tokens, so the context size matches its predecessor. The changes are the architecture, native vision and the price.

For background on the lineup, see our guide to Alibaba’s Qwen models and tools. Our report on a Qwen plugin for steering coding agents covers the team’s agent tooling.

What would confirm the efficiency claims

The announcement leaves several gaps. It gives no parameter counts, no number of active experts and no output limit. It also publishes no benchmark table in the post. Independent testing of token use and task cost against Qwen3.8-Max would show whether the 50% and 60% figures hold outside Alibaba’s internal setup.

Developers should also check whether the new prices apply to every context length and to the Token Plan. The post does not say. The promised small open-weight Qwen4 models are the other item to watch, since they would let teams run the family on their own hardware.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like