Qwen4 Max scores 43, $3.27 per task: how much are Kimi's 3 points still worth?
Qwen4 Max scores 43 on the Artificial Analysis Intelligence Index, three points behind Kimi K3.1 Pro. Its cost per task is $3.27, about 11% lower than Kimi and about 67% lower than Qwen3.8 Max. With scores and bills side by side, the choice is no longer just about who tops the leaderboard.
Artificial Analysis has updated its independent evaluation of Qwen4 Max. Qwen4 Max uses xhigh reasoning and scores 43 on the Intelligence Index, 9 points higher than Qwen3.8 Max and 3 points lower than Kimi K3.1 Pro. Its cost per task is $3.27, versus $3.68 for Kimi K3.1 Pro, about 11% lower; the previous generation Qwen3.8 Max was $10.02. The claims at launch differ from third-party results. Some online commenters noted that Alibaba said at launch that Qwen4 Max surpassed Kimi K3.1 Pro, but this index places it 3 points behind. That is also the point of third-party evaluations: beyond the launch table, you still need to look at reproduced results.
On sub-scores, Qwen4 Max's improvements over Qwen3.8 Max are very consistent: HLE-Diamond rises from 24% to 45%, up 21 points; AA-Omniscience up 18 points; AutomationBench-AA up 17 points; CritPt v2 up 13 points; Terminal-Bench-Science up 11 points; Terminal-Bench 4.0 up 9 points; GDP.pdf up 3 points; FrontierCode 2.0 up 3 points; AA-Briefcase improves by 82 Elo. AA-Briefcase covers agentic work in real business documents. Qwen4 Max reaches 1686 Elo, versus 1674 for Kimi K3.1 Pro and 1446 for Tencent Hy4. In knowledge-work scenarios, this score is clearly competitive.
But it does not lead across the board. On Terminal-Bench 4.0, Qwen4 Max is 48%, Kimi K3.1 Pro is 67%, and Tencent Hy4 is 37%. On FrontierCode 2.0, Qwen4 Max is 26%, tied with Kimi K3.1 Pro and below Tencent Hy4's 31%. If the core of the workflow is agentic coding in the terminal, this set of data is more worth looking at than the overall leaderboard ranking.
Cost reduction is not just from price cuts. Qwen4 Max uses about 105k output tokens per task, 39% fewer than Qwen3.8 Max's 171k; the input tokens used to run the index are 9.9B, versus 14.5B, 32% less. The cache-hit price is $0.075 per million tokens, versus $0.25 for the previous generation, 70% lower; the input price is $1.50 per million tokens, versus $2.00 for the previous generation, 25% lower. In agentic evaluations, the context is resent every round, and input accounts for about 85% to 90% of the cost of running the index; cache pricing and input tokens together brought down the per-task cost.
The model itself supports a 1M token context window, listed on the page as 984k, with text and image input and text output. First-party API pricing is $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 for cache hits. Tencent Hy4 has a lower cost per task, $1.03, but its index is 8 points lower. Between cheap and high-scoring, there is no free lunch.
Some online users have done the math: paying a 41-cent per-task premium for 3 points is acceptable for back-office tasks. Others remain skeptical about benchmark stability, arguing that results run by different people may not be consistent. For the full comparison, see Artificial Analysis's model page: https://artificialanalysis.ai/models/qwen4-max
When choosing a model, a 3-point difference in total score is only the entry point. If the task mix is documents, knowledge work, and automated workflows, Qwen4 Max's price and AA-Briefcase performance are more persuasive; if you run terminal coding every day, Kimi's 67% vs. 48% on Terminal-Bench 4.0 cannot be skipped. Leaderboards show position, bills show daily use, and the cost of a single failure must also be factored in.
发布时间: 2026-11-16 19:29