Across five custom business benchmarks reported by Spring Prompt on 2026-11-16, Qwen4-Max at the default high effort ranked near the top on DeckBench and SurveyBench, toward the lower middle on CatalogBench and ROASBench, and showed high latency on BulletBench. Some of the public tasks can be rerun. These results apply only to this model version, provider, and test setup.
Tasks this can help assess: Presentation generation, survey planning and analysis, product listing generation, simulated ad budget allocation, and time-limited chess decisions.
Tasks these results should not be generalized to: General reasoning, real commercial returns, coding, or long-running agents; this article did not test those capabilities.
Model version covered: Qwen4-Max; no API snapshot ID was provided.
Test environment or client: Spring Prompt's custom benchmarks; calls to qwen/qwen4-max through OpenRouter, with Alibaba serving the model.
Reasoning effort and parameters: No reasoning parameter was explicitly sent; the default effort was high on both OpenRouter and the Alibaba API. BulletBench also tested low; its best-results table uses low.
The results were published on 2026-11-16. Spring Prompt says it measured all results itself; each leaderboard ranks models separately, so scores on different scales should not be compared directly. DeckBench, SurveyBench, and ROASBench were each run once. On CatalogBench, each brief was repeated three times across 56 fictional products. On BulletBench, each configuration typically played 8–12 games.
| Benchmark | Tasks and sample | Scoring method |
|---|---|---|
| DeckBench | Create presentations from analytical materials; 6 tasks, 1 presentation per task | Pairwise comparisons based on numerical accuracy, narrative, programmatic layout checks, and slide-by-slide design review, converted to a rating; judges were GPT-6.1 Sol, Claude Opus 5.5, and Gemini 3.1 Pro, with majority vote |
| SurveyBench | Plan user surveys and analyze results; 9 tasks, 1 run per task | Checks survey quality, numerical analysis, detection of data traps, and recommendations; pairwise comparisons converted to a rating, with the same judging panel as DeckBench |
| CatalogBench | Generate listings for 56 fictional products from product feeds, images, and vendor copy; 3 rounds each for factual and sales briefs | Main metric: share of products that were publishable after all three rounds; unsupported claims, time, and cost were also counted. GPT-6.1 Sol judged the results, with a bias check |
| ROASBench | Simulate a brand's ad budget decisions over 12 consecutive months; 1 run per model | Budget allocation, adjustments based on the previous month's results, and whether the budget was exceeded, combined into a score out of 100 |
| BulletBench | Play against Stockfish under a real chess clock; 8–12 games per configuration | Evaluates Elo, move time, timeouts, and illegal moves; Blitz 3+2 uses default settings only |
Price comparisons use blended list price at an input-to-output token ratio of 3:1. Qwen4-Max costs $2.25 per million blended tokens; the "$1–3" tier refers to that blended price range. Task costs are estimates based on runtime prices and exclude evaluation costs.
| Benchmark and setting | Qwen4-Max result | Leaderboard position | Related cost or observation |
|---|---|---|---|
| DeckBench, high | Rating 1,585 (95% interval 1,497–1,708) | 1st of 32 models; 1st of 6 models in the $1–3 tier | About $0.18 and 8 minutes per presentation. Higher than GPT-6 Astra's 1,566, but their intervals overlap; none of the 6 presentations Qwen publicly showcased was ready to present, mainly because font sizes were below 12 pt |
| SurveyBench, high | Rating 1,395 (1,320–1,482); survey plan 95/100, analysis 89/100; handled 98.6% of data traps | 5th of 23 models; 1st of 4 models in its price tier | About $0.20 and 11.5 minutes per task; 3 of 9 public tasks had no errors, while analysis in 5 tasks included at least one conclusion unsupported by the data |
| CatalogBench, factual brief, high | 39.3% (22/56) publishable across all three rounds; 0.70 unsupported claims per product | 19th of 32 models; 3rd of 6 models in its price tier | About $0.007 and 21 seconds per product; GPT-6.1 Terra scored 64.3% |
| CatalogBench, sales brief, high | 8.9% publishable across all three rounds; 2.56 unsupported claims per product | 21st of 32 models; 3rd of 6 models in its price tier | Accuracy dropped markedly with sales-oriented wording |
| ROASBench, high | 45.8/100; simulated profit $775,520; did not exceed budget in any of the 12 months | 15th of 32 models; 2nd of 7 models in its price tier | About $0.69 per run; a single simulation does not represent real ad returns |
| BulletBench, Blitz 3+2, high | Rating 147; lost 7 of 8 games on time; median 8.6 seconds per move | 20th of 20 models | Three minutes plus 2 seconds per move on the clock; tests decisions under time pressure |
| BulletBench, 60 seconds, low | Rating 147 | 25th of 33 configurations; 4th of 6 configurations in its price tier | All of the first four games at high timed out, so testing stopped and no rating was assigned |
Overall ranks in the table reflect the snapshot in the review as of November 16. The corresponding live benchmark pages showed different model counts that day: DeckBench had 33 versus 32 in the review table, SurveyBench 24 versus 23, and ROASBench 33 versus 32. Qwen's rank was unchanged, but the model counts should not be treated as the same leaderboard snapshot.
Spring Prompt lists Qwen4-Max's combined leaderboard score as 56 points, 16th of 36, but on that date it covered only 2 of 5 benchmark groups; all data came from Spring Prompt's own benchmarks, and the overall rank was provisional.
For Qwen3.8 Max, the article reports higher Qwen4-Max results on every tested item: DeckBench 1,585 vs. 1,035; SurveyBench was newly tested; factual product listing rate 39.3% vs. 28.6%; ROASBench 45.8 vs. 25.4. Samples and run counts for individual benchmarks were limited.
Spring Prompt reports international endpoint API list prices of $1.50 for input and $4.50 per million tokens for output, with cached input at $0.075. These are the prices listed on the article's publication date and may change. Market price tiers use the 3:1 blended-price basis and cannot be compared directly with input-only or output-only prices.
Spring Prompt's public data page lists this release as 2026-11-16-49bdfa045dee and provides some benchmark results and public task data under a CC BY 4.0 license. The page notes that CatalogBench product images are not included in the download, and that ROASBench and BulletBench tasks are generated by a simulator or chess-game code rather than provided as task files.
These tests support a limited conclusion: Qwen4-Max scored near the top on Spring Prompt's one-off presentation and survey analysis tasks, while performance was weaker on product listings and tasks requiring quick responses. The increase in unsupported claims with sales briefs suggests that "more persuasive sales copy" may reduce factual reliability.
Most business benchmarks, except CatalogBench, had only a single run. Six presentations, nine surveys, and one ad simulation are not enough to reliably infer real-world business success rates. The rating intervals reflect uncertainty in pairwise comparison scores, not variance across repeated runs. Leaderboard positions are based on models included on that date and each benchmark's best settings, and will change over time; the price tiers also contain few models. The test data used fictional products or simulated markets and does not represent real product catalogs, surveys, or advertising outcomes.
The Spring Prompt Open Data page provides the release data and public tasks for DeckBench, SurveyBench, and CatalogBench, which can be rerun using the metrics on each benchmark page. The data page explicitly says the download excludes CatalogBench product images; ROASBench and BulletBench simulation tasks are not published as task data files, and model outputs are not yet included in the download. The public materials therefore make it possible to inspect the tasks and results, but are not enough to independently recreate every original run. Cite Spring Prompt, the release date, and CC BY 4.0 when using the data.
Qwen4-Max