Skip to content
Rush Commerce
AI & Automation3 min read

Qwen4-Max released: $1.50 input, cost per task down 62%

Alibaba released Qwen4-Max with a 1M-token context and API access at $1.50 input and $4.50 output. It is not open weights. Measure cost per task before you switch.

Qwen4-Max is out of preview, and the number to read is cost per task. Alibaba released Qwen4-Max on November 15 as its new flagship model. Four days ago we covered the preview that had an expiry date and no pay-as-you-go access. The release changes that: the model is now on the API. Alibaba says it does the same work as the last flagship with fewer tokens. That claim matters more than the token price.

What actually happened

The details come from the Qwen team's announcement:

  • The model. A new Qwen4 architecture with 2.4 trillion backbone parameters, of which 71 billion are active for each token. It has native vision and a one-million-token context window.
  • The access. Available now on QwenCloud through the Token Plan or the API. It is also in QwenWork, Qoder and QwenStudio.
  • The efficiency claim. On Terminal-Bench 4.0, Alibaba says token use fell about 48% to 49% against Qwen3.8-Max. On its own Work360 benchmark, token use fell 24% to 35%, and the average cost per task fell from $4.43 to $1.67. That is a 62% cut. Work360 is Alibaba's benchmark, so this is a vendor number.
  • The weights. Qwen4-Max is not an open-weight release. Alibaba says smaller Qwen4 models "are coming to open source." It gives no date.
  • The safety choice. Alibaba says it removed training data for offensive cyber tasks and kept data for defensive tasks such as vulnerability detection and remediation. It also says the share of code that passed both function and security checks rose from 22.0% to 31.8%. That still means most generated code failed one of the two checks.

Artificial Analysis gives the independent view:

  • Price. $1.50 per million input tokens and $4.50 per million output tokens, with a 95% cache discount.
  • Score. 43 on its Intelligence Index, rank 38 of 81 models. The median for comparable models is 41.
  • Cost to run. $3.27 per index task. The model produced 127 million output tokens in the test, against a median of 78.6 million. Artificial Analysis calls that "very verbose."
  • Context. It lists 984k tokens, a little under Alibaba's one million.

Why the Qwen4-Max release matters for your business

Two sources measured the same model and got different results. Alibaba says it uses fewer tokens. Artificial Analysis says it uses more tokens than most models. Both can be true, because each ran different tasks. Your tasks are a third set.

That is why the price per token is the wrong number to compare. A cheap model that writes long answers and retries can cost more than an expensive model that finishes in one pass.

What to do:

  1. Pick 20 real tasks from your own work. Quotes, support replies, code changes.
  2. Run them on your current model and on Qwen4-Max. Record total cost and whether the result was usable.
  3. Divide cost by usable results. That is your cost per task.
  4. Check where the data goes. This is a hosted model from Alibaba Cloud. Read the data terms before you send customer records.
  5. Do not plan on the open weights. There is no date.

Key takeaways

  • Alibaba released Qwen4-Max on November 15 with 2.4 trillion parameters, native vision and a one-million-token context
  • It is available on the API and the Token Plan; the weights are not open, and smaller open Qwen4 models have no date
  • Artificial Analysis lists $1.50 input and $4.50 output per million tokens, and a score of 43 on its Intelligence Index
  • Alibaba claims a 62% lower cost per task on its own benchmark; Artificial Analysis found the model very verbose
  • Measure cost per usable result on your own tasks before you change models

The cheapest token is not always the cheapest task. We build model routing with the cost of each task measured, so you can change vendors when the numbers change. Run your numbers in our ROI calculator, or see how we build systems that are not tied to one model.

Sources: Qwen, Artificial Analysis.

  • #qwen4-max
  • #alibaba
  • #model-pricing
  • #ai-agents
  • #cost-per-task
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.