Artificial Analysis reports that Qwen4 Max scored 43 on its Intelligence Index (xhigh), 9 points above Qwen3.8 Max (xhigh). The estimated cost per index task is $3.27, below Qwen3.8 Max's $10.02 and Kimi K3.1 Pro (max)'s $3.68, but Qwen4 Max used more output tokens than Kimi and scored lower on Terminal-Bench 4.0.
Suitable for: Comparing aggregate scores, sub-benchmark results, token usage, and estimated task costs in the same Artificial Analysis index snapshot.
Not suitable for extrapolating to: Other task sets, real-world business success rates, actual bills, or absolute rankings across different evaluation frameworks.
Model versions covered: The post's Qwen4 Max; comparisons are Qwen3.8 Max (xhigh), Kimi K3.1 Pro (max), and Tencent Hy4.
Test environment or client: Artificial Analysis Intelligence Index. The post does not disclose the full task inputs, prompts, API deployments, sample count, or complete run configuration.
Reasoning levels and parameters: Qwen4 Max and Qwen3.8 Max use xhigh; Kimi K3.1 Pro uses max. Other parameters are unspecified.
The Artificial Analysis model page identifies the Intelligence Index as v5.0.1, comprising nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1. The X post does not specify the index version, so the version and composition here come from the model page as viewed on 2026-11-17; they do not establish which version label applied when the post was published.
The post says Qwen4 Max scored higher than Qwen3.8 Max (xhigh) on all nine evaluations. On HLE-Diamond, the score rose from 24% to 45%; AA-Briefcase uses Elo ratings. Task cost and token usage are Artificial Analysis's aggregate measures for the index evaluations, not publicly itemized API bills for each question.
| Metric | Qwen4 Max | Comparison and difference |
|---|---|---|
| Intelligence Index | 43 (xhigh) | Qwen3.8 Max (xhigh) 34; Kimi K3.1 Pro (max) 46 |
| Cost per index task | $3.27 | Qwen3.8 Max $10.02; Kimi $3.68; Tencent Hy4 $1.03, with an index score 8 points lower |
| Output tokens per task | About 105k | Qwen3.8 Max about 171k (39% fewer); Kimi about 72k; Hy4 about 78k |
| Input tokens for the index run | About 9.9B | Qwen3.8 Max about 14.5B (32% fewer) |
| AA-Briefcase | 1686 Elo | Kimi 1674; Tencent Hy4 1446 |
| Terminal-Bench 4.0 | 48% | Kimi 67%; Tencent Hy4 37% |
| FrontierCode 2.0 | 26% | Kimi 26%; Tencent Hy4 31% |
Qwen4 Max's gains over Qwen3.8 Max on the other index components were: AA-Omniscience +18, AutomationBench-AA +17, CritPt v2 +13, Terminal-Bench-Science +11, GDP.pdf +3, and FrontierCode 2.0 +3; AA-Briefcase increased by 82 Elo. These are component-level differences reported in the post. Scores from different evaluations should not be added together or treated as the same kind of success rate.
Artificial Analysis reports a Qwen4 Max Intelligence Index score of 43, 9 points above Qwen3.8 Max (xhigh); Kimi K3.1 Pro (max) scored 46.
Cost per index task is a weighted estimate: $3.27 for Qwen4 Max, $10.02 for Qwen3.8 Max, $3.68 for Kimi, and $1.03 for Tencent Hy4. The post says Qwen4 Max's lower index run cost versus Qwen3.8 Max came from both lower prices and lower token usage; input accounted for about 85%–90% of the total cost for both.
The post lists Alibaba's official API prices per million input/output tokens: $1.50/$4.50 for Qwen4 Max, with cache hits at $0.075; and $2.00/$6.00 for Qwen3.8 Max, with cache hits at $0.25. These prices and costs reflect the terms reported when the post was published and may change later.
Component differences as recorded in the post: HLE-Diamond 24%→45% (+21 percentage points), AA-Omniscience +18, AutomationBench-AA +17, CritPt v2 +13, Terminal-Bench-Science +11, Terminal-Bench 4.0 +9, GDP.pdf +3, FrontierCode 2.0 +3, and AA-Briefcase +82 Elo.
In this Artificial Analysis index snapshot, Qwen4 Max scored higher than Qwen3.8 Max, with lower estimated costs and token usage; its aggregate score was still below Kimi K3.1 Pro's. It did not lead on every component: its Terminal-Bench 4.0 score was below Kimi's, and its FrontierCode 2.0 score was below Tencent Hy4's. Hy4 had a lower task cost but an aggregate score 8 points lower.
The index combines different tasks and scoring methods, so it should not be interpreted as a single task pass rate. The post and model page do not publish the full prompts, per-question data, model endpoint details, number of repetitions, or uncertainty intervals, so external readers cannot reproduce each result. Costs are estimates based on the listed prices and index usage, not bills for a particular account. The scores represent only this published snapshot and cannot establish performance across all tasks.
The original thread publishes aggregate scores, some component differences, token usage, and cost estimates, but not the full task set or execution configuration. Readers can consult the Artificial Analysis page for current model details and index composition. For a comparison, confirm that the index version, reasoning level, provider prices, and cache billing terms match, and interpret each evaluation separately.
Original X thread: https://x.com/ArtificialAnlys/status/2122161629592953032
Artificial Analysis model page: https://artificialanalysis.ai/models/qwen4-max
Qwen4-Max