Alibaba's Qwen4-Max Scores 43 on Artificial Analysis and Cuts Cost per Task to $3.27 From $10.02, but Trails Kimi K3.1 Pro by 19 Points on Terminal-Bench 4.0
Alibaba released Qwen4-Max on 16 November, a week after a preview build appeared on 11 November. Artificial Analysis scores it 43 on its Intelligence Index v5.0.1 at the highest reasoning setting, up 9 points from Qwen3.8-Max (34) and 3 points behind Moonshot’s Kimi K3.1 Pro (46). That puts it ninth among the ten leading models AA compares directly, and 38th of 81 in its model class, just above the class median of 41.
The more interesting number is cost. AA puts Qwen4-Max at $3.27 per Intelligence Index task, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro.
Where the cost drop comes from
- Output tokens per task fall to about 105,000, 39% below Qwen3.8-Max. Kimi K3.1 Pro still uses fewer, at about 72,000.
- List pricing is $1.50 per million input tokens and $4.50 per million output tokens, 25% below the predecessor on both. Cached input is $0.075 per million, a 70% cut.
- AA puts input tokens at roughly 85% to 90% of cost in multi-step agent tasks, because earlier context is resubmitted on each step. Cheaper cache hits therefore matter more than the output price.
The model is a mixture-of-experts with 2.4 trillion total parameters and 71 billion active per token, about 25% fewer active parameters than Qwen3.8-Max’s 95 billion at the same total size. A separate 328 billion parameters sit in an N-gram embedding table. Context is 1 million tokens, with text, image and video in and text out.
Benchmarks
| Benchmark | Qwen4-Max | Qwen3.8-Max | Kimi K3.1 Pro |
|---|---|---|---|
| AA Intelligence Index | 43 | 34 | 46 |
| Cost per AA task | $3.27 | $10.02 | $3.68 |
| Terminal-Bench 4.0 | ~48.5% | 38.9% | ~67% |
| FrontierCode 2.0-AA | 26% | n/a | 26% |
| AA-Briefcase (Elo) | 1,686 | ~1,605 | 1,674 |
| Vals Index | 63.20% | 59.05% | 62.52% |
Tencent’s Hy4 scores 35 on the index but beats Qwen4-Max on FrontierCode 2.0, 31% to 26%. AA reports Qwen4-Max improving on all nine evaluations in the index, with HLE-Diamond rising from 24% to 45%.
The coding gap is the weak spot. On Terminal-Bench 4.0, Kimi K3.1 Pro leads by roughly 19 points. Qwen4-Max is stronger on knowledge-work tasks: it edges Kimi on AA-Briefcase and the Vals Index, where it ranks 11th, the highest placement for a Chinese model on that list.
Arena results
Arena lists Qwen4-Max at 14th on Agent Arena with a +7.5% net improvement score, up from 39th for Qwen3.8-Max. On Code Arena WebDev it debuts sixth with a preliminary 1,750, 74 points above Qwen3.8-Max (1,676) and 66 behind Claude Opus 5.5 (1,816). Both are preliminary results that will move as votes accumulate.
What Alibaba claims versus what testers found
Alibaba’s launch post leans on Code Arena, Agents’ Last Exam, Vals Index, JobBench and Chartography, plus Work360, an in-house benchmark covering 721 professional fields. On Work360 it reports 35% fewer tokens at high effort and about 60% lower cost per task than Qwen3.8-Max. The independent cost figure is of similar size: AA’s $3.27 is about 67% below $10.02.
Alibaba’s own chart reports Qwen4-Max at 75.7 on DeepSWE v1.1, level with Opus 5.5 (74.2) and GPT-6.1 Sol (75.2), while conceding Terminal-Bench 4.0 at 48.5 and ProgramBench at 31.5, where Opus 5.5 scores 65.0. Alibaba’s results on professional and legal agent work are also uneven in independent testing: Vals measured Tax Agent Bench at 40.35% against 49.7% for Qwen3.8-Max, and Legal Research Bench at 41.35% against 46.2%.
Availability
Qwen4-Max is proprietary. It is served through QwenCloud’s API and Token Plan, QwenWork, Qwen Studio, Qoder, OpenRouter and Vercel AI Gateway. Alibaba says smaller Qwen4 models will be released with open weights, built on the same architecture and training methods, but has given no date.
The cost comparison with Western models remains unflattering on a per-intelligence basis. GPT-6.1 Sol scores 55 on the same index at $1.27 per task, and Gemini 4 Carbon scores 60 at $5.48. Qwen4-Max is a clear cost improvement within the Qwen line, not a challenge to the closed frontier.