I noted Alibaba’s Qwen4-Max launch as a case study in the trade-off between model capability and cost. The post emphasizes that the benchmark figures are company-reported, while the question of whether peak intelligence or lower cost matters more remains open.
Qwen4-Max arrived on 15 November.
Qwen4-Max has 2.4 trillion parameters; 71 billion active per token; a one-million-token context.
Qwen4-Max offers native vision; built for coding and long-running agents.
On Alibaba's own tests, Qwen4-Max leads on DeepSWE at 75.7 against Claude Opus 5.5 at 74.2.
On Terminal-Bench, Qwen4-Max scores 48.5 to 59.6 against Opus 5.5.
Artificial Analysis puts it at 43 on its Intelligence Index, against 59 for Opus 5.5.
At $1.50 per million input tokens, Alibaba says the cost per task fell from $4.43 to $1.67.
The model is not open weight; Max stays closed.
The headline numbers are Alibaba's own, and the post does not provide the task mix or assumptions behind its reported cost-per-task reduction.
Smaller Qwen4 models get open weights later, with no date given.
I leave open the question: Is the next race about the smartest model or the cheapest one that is good enough?