November 17, 2026AgentsAPIBenchmark

Qwen4-Max Sells on Cost Per Task, and Keeps the Weights

Alibaba finally shipped Qwen4-Max, and it was the top Hacker News story of the day: 846 points and more than 460 comments. The headline number is not a benchmark. It is a bill. On Work360, Alibaba's own agent benchmark, the average cost per task falls from $4.43 on Qwen3.8-Max to $1.67, a 62% cut. Blog: https://qwen.ai/blog?id=qwen4-max , pricing: https://www.qwencloud.com/models/qwen4-max

The specs, all from Alibaba: a 2.4 trillion parameter mixture-of-experts backbone with 71B active per token, plus 328B of N-gram embedding parameters, Gated DeltaNet mixed with Qwen Sparse Attention, native vision and a 1M-token context. Price is $1.5 in and $4.5 out per million tokens, with cached input at $0.075. Alibaba says the model uses roughly a quarter to a third fewer tokens on Work360 and about half as many on Terminal-Bench 4.0. That is where the savings come from: fewer tokens per job, not just a cheaper token.

Now read the chart closely. The comparison set is Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. Fable 5.2 and Gemini 4 are not on it at all. And even against that set, Qwen4-Max trails where agent builders care most: 48.5 on Terminal-Bench 4.0 against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1, and 31.5 on ProgramBench against Opus at 65.0. It wins on its own picks, like a legal benchmark and Agents' Last Exam, and edges DeepSWE by a hair. Arena puts it at #14 on Agent Arena, the top Chinese model there. So this is not a frontier crown. It is a credible, cheap, long-running worker.

The product idea is more interesting than the leaderboard. Alibaba calls it Live Delivery: ship a usable first version early, then keep releasing while the user can redirect. The plugin is open source under Apache-2.0 at https://github.com/QwenLM/Qwen-Live-Delivery-Plugin . The demos lean the same way, including a 48-hour autonomous run with a partner lab that ended with a physically synthesized coolant molecule, and a drug-discovery run of 454 tool calls that stayed computational.

The awkward part: the flagship is closed. No weights, API only through QwenCloud, Qoder and Alibaba's own apps, with smaller family members promised as open source later. The lab that made open weights its brand now keeps the best model behind a meter, exactly like the Americans it competes with. When the selling point becomes cost per task, the weights become the margin.
← Previous
Nerfed Daily: 2026-11-16
Next β†’
California Sues OpenAI Over Agents It Says Got Loose 80 Times
← Back to all articles

Comments

Loading...
>_