In the Vals Index v2.2, updated by Vals AI on 2026-11-15, Qwen4 Max scored 63.20% (±1.01) at the xhigh reasoning level, ranking 11th among 38 models; the industry-weighted index run cost $9.33. The score is not a general task success rate, and the cost is not a fixed price for a single task.
Tasks it can help assess: Compare performance on finance, coding, legal, and tax tasks, along with the GDP-weighted aggregate score, within the same Vals AI snapshot.
Tasks it should not be extrapolated to: General question answering, uncovered industries, real-world business performance, or absolute rankings across different evaluation frameworks.
Model version covered: Vals AI identifies it as Qwen 4 Max; the model page's publication date is 2026-11-15. Alibaba's press release was published on 2026-11-16; the dates use different conventions.
Test environment or client: Vals Index v2.2. The page says it combines private and public benchmarks; it does not disclose item-level inputs, the full harness, or run logs.
Reasoning level and parameters: xhigh; temperature 1.0, top_p 0.95, top_k 20. The model page lists a 1M context window and a 262K maximum output.
The Vals Index weights scores across four industries by their estimated shares of U.S. GDP: finance 8.0, coding 5.6, legal 1.2, and tax 0.5, for a total weight of 15.3. It first averages the benchmarks within each industry, then aggregates them as follows:
金融 = AVG(Finance Agent v2.1, EMB v1.1)
编码 = AVG(Terminal-Bench 4.0, Vibe Code Bench v1.1, Code Migration v1.1)
法律 = AVG(Legal Research Bench, HLAB-Vals-1.1)
税务 = Tax Agent Bench
Vals Index = (8.0×金融 + 5.6×编码 + 1.2×法律 + 0.5×税务) / 15.3The benchmarks cover multi-step financial reasoning, spreadsheet modeling, command-line and application building, code migration, legal research citing legal materials and long-form work products, and U.S. corporate tax question answering with supporting evidence. Vals estimates weights using industry value-added data published by the BEA on FRED, and estimates potential economic impact from task performance. Vals Index v2.2 methodology
The following results were shown on the model page when accessed on 2026-11-17; uncertainty intervals are ±1 standard error.
| Metric | Qwen4 Max | Page rank |
|---|---|---|
| Vals Index v2.2 | 63.20% ±1.01 | 11/38 |
| Vals Finance Index | 61.99% ±1.45 | 8/41 |
| Vals Legal Index | 70.71% ±1.75 | 17/39 |
| Finance Agent v2.1 | 66.68% ±2.02 | 8/41 |
| EMB v1.1 | 70.83% ±2.36 | 18/43 |
| Terminal-Bench 4.0 | 48.48% ±0.88 | 15/65 |
| Vibe Code Bench v1.1 | 86.66% ±2.27 | 15/129 |
| Code Migration v1.1 | 57.3% | Rank and uncertainty interval not listed on the page |
| Legal Research Bench | 41.35% ±3.42 | 33/95 |
| HLAB-Vals-1.1 | 21.25% ±3.33 | 7/39 |
| Tax Agent Bench | 40.35% ±3.54 | 13/86 |
The Vals Index leaderboard also shows that the run used about 27.27M tokens, took about 1 hour 5 minutes, and cost $9.33; the model page lists a duration of 64 minutes 39 seconds. The Vals Index announcement says Qwen4 Max scored above Kimi K3.1 Pro at 62.52% and Qwen 3.8 Max 0902 at 59.05%.
The Vals model page lists token prices of $1.50/$4.50 per million input/output tokens; Alibaba Cloud's Model Studio documentation for the Singapore region lists the same prices for qwen4-max (international deployment, input length up to 1M). These are vendor list prices, not the cost of a complete Vals Index run.
$9.33 per run is the evaluation cost on the Vals Index leaderboard and varies with evaluation token usage and the input/output mix. The page notes that the index cost includes Code Migration; public materials do not break down that run's bill by caching, region, or discounts.
The Vals model page says the model weights are private. Its index methodology discloses the benchmark composition, industry weights, and aggregation formula, but some industry benchmarks use private question sets and do not provide all questions or item-level results.
Date conventions: the Vals model page gives the publication date as 2026-11-15, and the Vals Index page is also marked as updated on 2026-11-15; Alibaba's press release was published on 2026-11-16. The dates from each source are retained here rather than combined into a single publication date.
This snapshot places Qwen4 Max near the top of the Vals Index, with stronger component scores in finance and application building and a lower score in legal research; the component leaderboards and overall leaderboard have their own sample sizes and ranking ranges. The GDP-weighted aggregate reflects performance on the benchmarks and weights selected by Vals. It cannot be interpreted to mean that “63.20% of economic work can be done by the model.”
The cost of a single test depends on usage, output length, caching, service region, and billing discounts. Official list prices can be used for estimates, but they do not replace the $9.33 run cost reported by Vals or directly represent an actual account bill. Private question sets, unpublished item-level materials, and the model-serving snapshot limit external reproducibility. Comparisons with scores from other organizations should account for each one's question sets, scoring methods, reasoning levels, and cost conventions.
The public Vals Index v2.2 formula can be used to understand the GDP weights and industry aggregation, and the public model page provides scores, rankings, and parameters. Because some benchmarks are private to Vals and the model weights are also undisclosed, external readers cannot independently rerun the complete index. The Alibaba Cloud unit prices are values from the Singapore-region page accessed on 2026-11-17; prices or service configurations may change later.
Vals AI model page: Qwen 4 Max Benchmarks, Cost and Capabilities
Vals AI index methodology: Vals Index v2.2
Alibaba Cloud Model Studio: Qwen4-Max pricing and model parameters
Alibaba press release: Alibaba Unveils Qwen4-Max
Qwen4-Max