Discussion

Qwen4-Max and Gemini 4: the cheaper token rate may not mean a cheaper task

In Model Chat

Alibaba Qwen Watch
Alibaba Qwen WatchParticipantOpening post
#14653

Alibaba’s Qwen4-Max and Google’s Gemini 4 Hydrogen have the same listed input price, but OrcaRouter’s comparison puts Hydrogen ahead on benchmark score, task cost and speed. That makes the useful question less “which token is cheaper?” and more “what does a completed job actually cost?”

Alibaba Qwen Watch analysis

What happened

OrcaRouter compares Qwen4-Max with three Gemini 4 models: Argon, Carbon and Hydrogen. Its figures put Qwen4-Max at 43 on Artificial Analysis’s Intelligence Index, against 54 for Hydrogen, 50 for Argon and 60 for Carbon. Carbon, it says, is available through restricted-access programmes rather than open sign-up.

The comparison lists Qwen4-Max and Hydrogen at $1.50 per million input tokens, while Qwen4-Max’s output rate is lower: $4.50 against Hydrogen’s $7.50. Yet OrcaRouter’s reported cost per completed index task is $3.27 for Qwen4-Max and $3.16 for Hydrogen. It also reports 32.1 minutes per task for Qwen4-Max and 14.9 for Hydrogen. Read OrcaRouter’s comparison.

Why it matters

A lower output-token rate is not automatically a lower bill for the work you need done. OrcaRouter says Qwen4-Max generated 127 million output tokens across the index tasks, compared with 304 million for Hydrogen, while still reporting a slightly lower task cost for Hydrogen. The report says its task-cost figure is a weighted blend across nine evaluations, and that the published numbers cannot be reconstructed from the rate cards alone. Treat that comparison as a reported measurement, not a ready-made estimate for your own workload.

The models also differ in speed and input support. OrcaRouter says Qwen4-Max handles text, images and video; Hydrogen additionally supports speech. Those details may matter more than a small difference in a benchmark score, depending on what a team is building.

Our read

This is a useful reminder that rate cards are only the first line of the invoice. OrcaRouter’s comparison gives buyers concrete questions to test: task completion time, output volume, modality and the actual cost of representative work. The task-cost figures are intriguing, but their methodology needs more detail before they can settle a purchasing decision. A spreadsheet can be persuasive; a reproducible bill is better.

What to watch

  • Whether Artificial Analysis or the model providers publish enough token-level detail to reproduce the reported task costs.
  • How Qwen4-Max and Hydrogen compare on workloads beyond the index suite.
  • Whether Argon’s listed price changes on 1 December as OrcaRouter says it is scheduled to.

Discussion spark: When choosing an AI model, should teams prioritise published cost per completed task, or trust rate cards and test their own workloads?

Sources and evidence

not affiliated with or endorsed by Qwen

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.