- Work360 (Xhigh effort) 52.7
- Agentic coding, DeepSWE v1.1 75.7
- Web development, CodeArena (Elo), cited from Arena 1750
- Agentic terminal coding, Terminal-Bench 4.0 (Xhigh effort) 48.5
- Black-box software replication, ProgramBench 31.5
- Agentic knowledge work, AA-Briefcase v1.1 (Elo), cited from Artificial Analysis 1686
- Long-horizon valuable tasks, Agents' Last Exam v1.1, cited from its leaderboard 46.7
- Professional job tasks, JobBench (updated dataset and evaluation code) 68.5
- Legal agent work, HLAB-Vals-1.1, cited from the Vals Index 21.3
- Agentic SaaS workflow, Automationbench-AA, cited from Artificial Analysis 72.5
- Real-world agent tasks, AgentArena (Net Improvement), cited from Arena 7.5
- Bioinformatics research, BioMysteryBench 85
- Clinical reasoning, HealthBench Professional 60
- Multidisciplinary reasoning, Humanity's Last Exam 51.7
- Visual chart understanding, Chartography (min_pixels 4096×32×32) 54.2
- Video understanding, Video-Index 68.1
- Last Secure Code Bench, code reconstruction without a security hint, function-level pass rate for both function and security 31.8
Qwen4-Max
Qwen4-Max ranks first on 2 of the 40 benchmarks with five or more models scored.
Qwen4-Max
Closed weights Flagship
- Vals Index v2.2 63.2% #14 /41
- Legal Research Bench 41.4% #37 /97
- Tax Agent Bench 40.4% #13 /87
How does it do against its rivals?
Qwen4-Max beats GPT-6 Sol on 12 of the 20 benchmarks they share. Muse Wonder does best against it, ahead on 15 of 30.
- GPT-6 Sol 12 ahead, 7 behind, 1 tied Compare
-
Muse Ember 11 ahead, 7 behind, 1 tied Compare
- GPT-6.1 Terra 17 ahead, 11 behind, 1 tied Compare
- Claude Fable 5 9 ahead, 7 behind Compare
- Claude Opus 5 10 ahead, 10 behind Compare
-
Muse Wonder 14 ahead, 15 behind, 1 tied Compare
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Caliper Bench v3 | 88.9 | #1 of 71 | Official leaderboard | |
| BullshitBench V2 | 90% | #6 of 137 | Official leaderboard | |
| SuperCLUE-LongContext | 51.4 | #1 of 13 | Official leaderboard | |
| LLM Debate Benchmark | 1,679 | #6 of 58 | Official leaderboard | |
| Extended NYT Connections | 90.8% | #26 of 109 | Official leaderboard | |
| Humanity's Last Exam | 51.7% | Not ranked | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BioMysteryBench | 85% | Not ranked | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Code Arena | 1,750 | #6 of 131 | Official leaderboard | |
| DesignArena Agentic Overall | 1,357 | #5 of 80 | Official leaderboard | |
| Image-to-WebDev Arena | 1,674 | #6 of 55 | Official leaderboard | |
| ALE-Bench | 2,342 | #15 of 134 | Official leaderboard | |
| Vibe Code Bench | 86.7% | #18 of 124 | Official leaderboard | |
| FrontierCS 2 | 45.8% | #6 of 20 | Official leaderboard | |
| RSI-Exam | 44.3% | #6 of 20 | Official leaderboard | |
| DAXBench v2 | 81.8% | #24 of 49 | Official leaderboard | |
| DeepSWE | 75.7% | Not ranked | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Agent Arena | 7.46% | #12 of 58 | Official leaderboard | |
| Toloka Arena Enterprise Tool Use (November 2026) | 65% | #12 of 43 | Official leaderboard | |
| PokéBench: Tool Assisted Benchmark | 10.0 | #18 of 34 | Official leaderboard | |
| KradleBench | 11.0 | #25 of 26 | Official leaderboard | |
| AutomationBench | 72.5% | Not ranked | Lab model card | |
| Terminal-Bench 4.0 | 48.5% | Not ranked | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FoodTruck Bench | $71,644 | #4 of 65 | Official leaderboard | |
| YC-Bench | $2,487,230 | #7 of 66 | Official leaderboard | |
| DesignArena Python-PPTX Slides | 1,302 | #4 of 37 | Official leaderboard | |
| DesignArena Spreadsheet | 1,337 | #5 of 40 | Official leaderboard | |
| Tax Agent Bench | 40.4% | #13 of 87 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 21.3% | #7 of 42 | Official leaderboard | |
| Vals Finance Agent v2.1 | 66.7% | #9 of 44 | Official leaderboard | |
| Vals Finance Index | 62% | #9 of 44 | Official leaderboard | |
| Vals Index v2.2 | 63.2% | #14 of 41 | Official leaderboard | |
| Legal Research Bench | 41.4% | #37 of 97 | Official leaderboard | |
| Vals Legal Index | 70.7% | #18 of 42 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 70.8% | #21 of 46 | Official leaderboard | |
| DAYJOB: Accounting | 22.4% | #13 of 26 | Official leaderboard | |
| DualEntry Accounting AI Benchmark | 83.2% | #15 of 29 | Official leaderboard | |
| DAYJOB: Law | 17.5% | #12 of 22 | Official leaderboard | |
| Deta Studio AI Construction Takeoff | 45 | #11 of 17 | Official leaderboard | |
| DataAnalyticsBench | 35.0 | #37 of 51 | Official leaderboard | |
| AA-Briefcase v1.1 | 1,686 | #4 of 10 lab numbers | Lab model card | |
| Work360 | 52.7 | #1 of 1 lab number | Lab model card | |
| Agents' Last Exam v1.1 | 46.7% | Not ranked | Lab model card | |
| HealthBench Professional | 60% | Not ranked | Lab model card | |
| JobBench | 68.5% | Not ranked | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BuseyBench SVG Test | 6.50 | #19 of 261 | Official leaderboard | |
| Chartography | 75.3% | #19 of 56 | Official leaderboard | |
| Video-Index | 68.1% | #1 of 1 lab number | Lab model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Last Secure Code Benchmark | 31.8% | Not ranked | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SuperCLUE | 61.6 | #3 of 18 | Official leaderboard | |
| Artificial Analysis Intelligence Index v5 | 42.9 | #19 of 69 | Official leaderboard | |
| LiveBench | 73.1% | #14 of 36 | Official leaderboard |
What does it cost to run?
On the 25 leaderboards that publish a price, its best run usually costs 1.2 times the median model's. On 2 of them, no model comes close to its score for less.
| Benchmark | Score | Price | From cheapest to priciest | Against the median | Close to its score for less Close for less |
|---|---|---|---|---|---|
| BuseyBench SVG Test | 6.50 | $0.23 18× the median | 18× | GPT-6 Luna 6.40 for $0.015 | |
| ALE-Bench | 2,342 | $2.56 3.1× the median | 3.1× | GPT-6.1 Terra 2,609 for $0.41 | |
| BullshitBench V2 | 90% | $0.013 2.6× the median | 2.6× | | |
| Vibe Code Bench | 86.7% | $41.40 10× the median | 10× | | |
| Legal Research Bench | 41.4% | $2.59 2.4× the median | 2.4× | GLM Legal 41.8% for $0.05 | |
| Tax Agent Bench | 40.4% | $1.03 1.4× the median | 1.4× | | |
| FoodTruck Bench | $71,644 | $7.15 per run 4.0× the median | 4.0× | None | |
| DataAnalyticsBench | 35.0 | $2.29 per run 0.7× the median | 0.7× | | |
| DesignArena Agentic Overall | 1,357 | $1.58 1.4× the median | 1.4× | GPT-6.1 Sol 1,360 for $1.17 | |
| Excel Modeling Benchmark v1.1 | 70.8% | $1.00 0.6× the median | 0.6× | | |
| Toloka Arena Enterprise Tool Use (November 2026) | 65% | $2.20 1.0× the median | 1.0× | | |
| Vals Finance Agent v2.1 | 66.7% | $1.74 1.1× the median | 1.1× | | |
| Vals Finance Index | 62% | $1.26 0.9× the median | 0.9× | | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 21.3% | $12.87 3.9× the median | 3.9× | | |
| Vals Legal Index | 70.7% | $5.50 2.3× the median | 2.3× | MiMo-V2.6 Flash 69% for $0.07 | |
| Vals Index v2.2 | 63.2% | $9.33 2.5× the median | 2.5× | | |
| DesignArena Spreadsheet | 1,337 | $0.77 1.1× the median | 1.1× | | |
| LiveBench | 73.1% | $0.51 0.6× the median | 0.6× | GPT-6.2 Sol 80.3% for $0.42 | |
| DesignArena Python-PPTX Slides | 1,302 | $2.88 2.6× the median | 2.6× | Kimi K3.1 Pro 1,286 for $0.9 | |
| DAYJOB: Accounting | 22.4% | $1.71 0.5× the median | 0.5× | GPT-6.2 Sol 27.6% for $1.51 | |
| KradleBench | 11.0 | $0.16 0.5× the median | 0.5× | GPT-6 Luna 18.1 for $0.021 | |
| DAYJOB: Law | 17.5% | $1.49 0.5× the median | 0.5× | GPT-6.2 Sol 32.7% for $1.42 | |
| Deta Studio AI Construction Takeoff | 45 | $1.55 1.2× the median | 1.2× | GPT-6.1 Terra 73 for $0.28 | |
| LLM Debate Benchmark | 1,679 | $0.071 0.5× the median | 0.5× | None | |
| Extended NYT Connections | 90.8% | $0.028 1.0× the median | 1.0× | GPT-6.2 Sol 94.1% for $0.0051 |
Inside agents
Boards that list Qwen4-Max inside an agent or product, such as Qwen Code + Qwen4-Max. Those scores belong to the agent, so we don't count them as the model's own. Its own score on the same board is beside each one.
| Benchmark | Entry | Agent's score | Its own score |
|---|---|---|---|
| PostTrainBench 1.2 | Qwen Code + Qwen4-Max | 36.7% |
Its lab's numbers put it at 164.7, number 20 of 412
The Self-Reported Index fits one score to the 8 capability results Qwen reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 159.7 to 169.7. How we compute it.
- BioMysteryBench 85
- DeepSWE 75.7
- AutomationBench 72.5
- JobBench 68.5
- HealthBench Professional 60
- Chartography 54.2
- Humanity's Last Exam 51.7
- Terminal-Bench 4.0 48.5
Who measured it?
Independent runs cover 39 of its 51 benchmarks. For the other 12, we have only Qwen's own number or a claim Epoch AI collected.
- An independent run and a claim 4
- An independent run only 35
- A claim only 12
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Its lab's numbers next to 4 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| Chartography | 54.2% Visual chart understanding, Chartography (min_pixels 4096×32×32) | 75.3% Leaderboard, xhigh | −21 pts |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 21.3% Legal agent work, HLAB-Vals-1.1, cited from the Vals Index | 21.3% Leaderboard | Same |
| Agent Arena | 7.50% Real-world agent tasks, AgentArena (Net Improvement), cited from Arena | 7.46% Leaderboard, xhigh | Same |
| Code Arena | 1,750 Web development, CodeArena (Elo), cited from Arena | 1,750 Leaderboard, xhigh | Same |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.