Grok 4 ranks first on 2 of the 115 benchmarks it has scores on.
Closed weights
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.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Fiction.LiveBench | 94.4% | #3 of 57 | Epoch AI | |
| MMLU-Pro | 87% | #18 of 212 | Official leaderboard | |
| SuperGPQA | 63% | #6 of 58 | Official leaderboard | |
| LLM Confabulation Leaderboard | 12.4% | #9 of 58 | Official leaderboard | |
| SimpleQA Verified | 51.9% | #17 of 56 | Official leaderboard | |
| FACTS Parametric | 58.6% | #14 of 45 | Official leaderboard | |
| FACTS Search | 75.3% | #10 of 31 | Official leaderboard | |
| LMArena (Elo) | 1,411 | #117 of 333 | Official leaderboard | |
| WritingBench | 74.7 | #18 of 45 | Official leaderboard | |
| Vectara Hallucination Leaderboard (HHEM) | 4.80% | #60 of 148 | Official leaderboard | |
| Humanity's Last Exam | 38.6% | #61 of 147 | Model card | |
| IndicGenBench | 52.0 | #7 of 13 | Official leaderboard | |
| Lech Mazur Creative Writing | 7.69 | #24 of 44 | Epoch AI | |
| Global MMLU Lite | 88.8% | #24 of 43 | Official leaderboard | |
| SkateBench v2 | 61.3% | #12 of 21 | Official leaderboard | |
| EQ-Bench 3 | 1,154 | #47 of 78 | Official leaderboard | |
| Phare | 62.1% | #46 of 62 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 45.7 | #92 of 119 | Official leaderboard | |
| FACTS Multimodal | 25.7% | #32 of 39 | Official leaderboard | |
| FACTS Benchmark Suite | 25.7% | #38 of 40 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2025 | 98.8% | #12 of 143 | Model card | |
| PHYBench | 42.3% | #3 of 28 | Official leaderboard | |
| HiPhO | 64.4% | #7 of 35 | Official leaderboard | |
| GPQA Diamond | 87% | #63 of 219 | Epoch AI | |
| HMMT 2025 | 93.9% | #15 of 49 | Model card | |
| BRUMO 2025 | 95% | #13 of 42 | Official leaderboard | |
| CMIMC 2025 | 85% | #12 of 33 | Official leaderboard | |
| SimpleBench | 60.5% | #37 of 100 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 84% | #82 of 202 | Epoch AI | |
| SMT 2025 | 85.8% | #17 of 40 | Official leaderboard | |
| FrontierMath | 19.7% | #31 of 70 | Epoch AI | |
| ARC-AGI | 79.6% | #46 of 92 | Official leaderboard | |
| ARC-AGI-2 | 29.4% | #48 of 95 | Official leaderboard | |
| ATLAS (OpenCompass) | 34.1% | #3 of 5 | Official leaderboard | |
| MathArena (Apex, IMO, USAMO) | 2.08% | #29 of 47 | Official leaderboard | |
| IMProofBench | 11.1% | #7 of 10 | Official leaderboard | |
| FrontierMath Tier 4 | 2.08% | #40 of 55 | Epoch AI | |
| HMMT Nov 2025 | 88.3% | #16 of 22 | Official leaderboard | |
| ForecastBench | 64.8 | #71 of 86 | Official leaderboard | |
| IMO 2025 | 11.9% | #5 of 6 | Official leaderboard | |
| xbench-ScienceQA | 65% | #28 of 29 | Official leaderboard | |
| USAMO 2025 | 37.5% | #2 of 2 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Aider Polyglot | 79.6% | #5 of 54 | Epoch AI | |
| AutoCodeBench | 50.9% | #4 of 34 | Official leaderboard | |
| METR Time Horizon | 105 min | #6 of 31 | Official leaderboard | |
| LiveCodeBench | 79.3% | #16 of 73 | Model card | |
| IOI (Vals) | 26.2% | #17 of 55 | Official leaderboard | |
| WeirdML | 45.7% | #59 of 112 | Official leaderboard | |
| FrontierCS | 24.0 | #11 of 14 | Official leaderboard | |
| OTelBench | 4.35% | #13 of 14 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BFCL V4 | 63% | #8 of 65 | Official leaderboard | |
| Game Arena Poker | 16.3 | #5 of 31 | Official leaderboard | |
| MCP-Universe | 33.3% | #5 of 30 | Official leaderboard | |
| Elimination Game | 5 | #12 of 56 | Official leaderboard | |
| LisanBench | 4042.2 | #29 of 133 | Official leaderboard | |
| Step Race | 3 | #13 of 59 | Official leaderboard | |
| MCPMark | 31.7% | #7 of 28 | Official leaderboard | |
| CompileBench | 86.7% | #5 of 20 | Official leaderboard | |
| BALROG | 43.6% | #9 of 35 | Official leaderboard | |
| BalatroBench | 9.47 | #6 of 16 | Official leaderboard | |
| Game Arena Chess Openings | 739 | #5 of 11 | Official leaderboard | |
| DeepResearch Bench | 47.3% | #11 of 24 | Epoch AI | |
| Galileo Agent Leaderboard v2 | 42% | #11 of 21 | Official leaderboard | |
| Game Arena Chess | 755 | #21 of 39 | Official leaderboard | |
| Meta Agents Research Environments (ARE) / Gaia2 | 15.7% | #7 of 13 | Official leaderboard | |
| Terminal-Bench | 27.2% | #10 of 18 | Epoch AI | |
| Game Arena Five in a Row | 442 | #12 of 20 | Official leaderboard | |
| Game Arena Four in a Row | 265 | #25 of 38 | Official leaderboard | |
| Toolathlon | 27.5% | #34 of 50 | Official leaderboard | |
| Terminal-Bench 2.0 | 27.2% | #34 of 43 | Official leaderboard | |
| Kaggle Game Arena | 181 | #31 of 39 | Official leaderboard | |
| Game Arena Word Association | 0 | #36 of 36 | Official leaderboard | |
| LMArena Search Arena | 1,123 | #35 of 35 | Official leaderboard | |
| Game Arena Werewolf | -6 | #8 of 8 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vending-Bench | $4,694 | #1 of 9 | Official leaderboard | |
| CaseLaw v2 | 65.8% | #9 of 53 | Official leaderboard | |
| CorpFin v2 | 66% | #24 of 119 | Official leaderboard | |
| Vals Finance Agent v1 | 40.3% | #18 of 47 | Official leaderboard | |
| Vals Finance Agent | 53.5% | #19 of 49 | Official leaderboard | |
| LegalBench | 83.2% | #57 of 138 | Official leaderboard | |
| ProfBench | 45.1% | #43 of 72 | Official leaderboard | |
| MedScribe | 78.2% | #61 of 98 | Official leaderboard | |
| MedCode | 38.1% | #67 of 96 | Official leaderboard | |
| prinzbench | 23.0 | #20 of 26 | Official leaderboard | |
| TaxEval v2 | 65.1% | #106 of 129 | Official leaderboard | |
| MortgageTax | 44.5% | #78 of 86 | Official leaderboard | |
| GDPval | 21.1% | #10 of 11 | Epoch AI |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| OCRBench v2 | 45% | #28 of 65 | Official leaderboard | |
| Blueprint-Bench | 0.39 | #4 of 8 | Official leaderboard | |
| ZeroBench | 6.00% | #39 of 66 | Official leaderboard | |
| Vision Arena | 1,184 | #89 of 136 | Official leaderboard | |
| BabyVision | 16.2% | #23 of 27 | Official leaderboard | |
| Math Kangaroo 2025 (MathArena) | 70% | #19 of 22 | Official leaderboard | |
| GeoBench | 45% | #44 of 48 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| General refusals (xAI) | 0% | #1 of 7 | Model card | |
| Virology Capabilities Test (VCT) | 60% | #5 of 28 | Model card | |
| BioLP-bench | 47% | #2 of 9 | Model card | |
| BaxBench | 43.6% | #8 of 34 | Official leaderboard | |
| Refusals + User Jailbreak (xAI) | 0% | #1 of 4 | Model card | |
| Cybench | 43% | #7 of 24 | Official leaderboard | |
| SycophancyEval (answer sycophancy) | 7.00% | #1 of 3 | Model card | |
| AgentDojo | 2.00% | #3 of 7 | Model card | |
| Soft Bias (Internal) | 0.36 | #1 of 2 | Model card | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 49.7% | #44 of 71 | Official leaderboard | |
| AgentHarm | 14% | #4 of 6 | Model card | |
| PCT (Polarized Contrastive Pairs) | 49.5% | #32 of 46 | Official leaderboard | |
| Refusals + System Jailbreak (xAI) | 1.00% | #3 of 4 | Model card | |
| MakeMeSay | 12% | #7 of 9 | Model card | |
| AIR-Bench 2024 | 44.4% | #70 of 82 | Official leaderboard | |
| MASK | 43% | #18 of 21 | Model card | |
| BinaryAudit | 8.77% | #18 of 20 | Official leaderboard | |
| Anthropic Red Team | 95.7% | #68 of 75 | Official leaderboard | |
| HELM Safety | 83.6 | #68 of 75 | Official leaderboard | |
| SimpleSafetyTests | 92.2% | #69 of 75 | Official leaderboard | |
| KORA | 14.8% | #39 of 42 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HELM Capabilities | 78.5 | #7 of 60 | Official leaderboard | |
| CAIS Text Capabilities Index | 26.2 | #41 of 66 | Official leaderboard | |
| CAIS Risk Index | 59.3 | #48 of 65 | Official leaderboard | |
| CAIS Vision Capabilities Index | 42.5 | #40 of 53 | Official leaderboard | |
| SAGE (Vals) | 25.1% | #79 of 81 | Official leaderboard |
xAI's models in color, against every other model on the Epoch Capabilities Index, by release date.
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 13 of them aren't in the launch post.