Introducing GPT-5.4 5 Mar 2026
GPT-5.4
GPT-5.4 ranks first on 8 of the 197 benchmarks it has scores on.
202
Benchmarks with a score
8
Ranked first, out of five or more models
35
In the top three
OpenAI Mar 2026
GPT-5.4
Closed weights Flagship
157.1
#25 of 271
Epoch Capabilities Index
- Humanity's Last Exam 36.2% #9 /47
- SWE-bench Verified 76.9% #8 /32
- SWE-Bench Pro 59.1% #2 /24
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 | |
|---|---|---|---|---|
| CL-bench | 27.9% | #1 of 24 | Official leaderboard | |
| MMLU-Pro | 87.5% | #9 of 212 | Official leaderboard | |
| HELM Arabic Enterprise | 76.7 | #2 of 23 | Official leaderboard | |
| SkateBench v2 | 81.5% | #2 of 21 | Official leaderboard | |
| LMArena (Elo) | 1,475 | #34 of 333 | Official leaderboard | |
| HELM Arabic | 82.5 | #3 of 28 | Official leaderboard | |
| Wer wird Millionär? LLM Benchmark | 275000.0 | #9 of 76 | Official leaderboard | |
| CL-bench Life | 21.7% | #2 of 16 | Official leaderboard | |
| Creative Writing v3 | 78.8 | #18 of 120 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 78.3 | #19 of 119 | Official leaderboard | |
| SEA-HELM | 76.0 | #10 of 62 | Official leaderboard | |
| Judgemark v2.1 | 85.0 | #7 of 42 | Official leaderboard | |
| EQ-Bench 3 | 1,537 | #14 of 78 | Official leaderboard | |
| MultiNRC | 58.3% | #7 of 38 | Official leaderboard | |
| Extended NYT Connections | 91.3% | #18 of 97 | Official leaderboard | |
| Humanity's Last Exam | 36.2% | #9 of 47 | Official leaderboard | |
| EQ-Bench 4 | 1,258 | #10 of 32 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 7.00% | #33 of 104 | Official leaderboard | |
| Judgemark v4 | 72.1 | #14 of 40 | Official leaderboard | |
| Document Arena | 1,470 | #14 of 37 | Official leaderboard | |
| BullshitBench V2 | 48% | #48 of 122 | Official leaderboard | |
| FACTS Multimodal | 41% | #17 of 39 | Official leaderboard | |
| BullshitBench | 25.4% | #59 of 107 | Official leaderboard | |
| SimpleQA Verified | 30.4% | #31 of 56 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FrontierMath | 47.6% | #4 of 70 | Epoch AI | |
| HMMT 2026 | 97.7% | #2 of 31 | Official leaderboard | |
| GPQA Diamond | 93.3% | #19 of 219 | Epoch AI | |
| AIME 2026 | 99.2% | #3 of 31 | Official leaderboard | |
| IMProofBench | 49.3% | #1 of 10 | Official leaderboard | |
| FrontierMath Tier 4 | 27.1% | #7 of 55 | Epoch AI | |
| MathArena (Apex, IMO, USAMO) | 54.2% | #6 of 47 | Official leaderboard | |
| CritPt | 23.4% | #23 of 160 | Epoch AI | |
| ProofBench | 56% | #9 of 59 | Official leaderboard | |
| DTBench | 90.7% | #28 of 178 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 97.8% | #32 of 202 | Epoch AI | |
| ARC-AGI-2 | 74% | #21 of 95 | Official leaderboard | |
| USAMO 2026 | 95.2% | #2 of 9 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 78.6% | #20 of 89 | Epoch AI | |
| ARC-AGI | 93.7% | #24 of 92 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 49% | #21 of 71 | Epoch AI | |
| ResearchClawBench | 18.8 | #10 of 23 | Official leaderboard | |
| Project Euler (MathArena) | 89% | #2 of 4 | Official leaderboard | |
| ForecastBench | 63.0 | #46 of 86 | Official leaderboard | |
| Leipzig Benchmark | 28% | #11 of 19 | Official leaderboard | |
| xbench-ScienceQA | 70.8% | #17 of 29 | Official leaderboard | |
| Echo General AI Prediction Leaderboard | 999 | #10 of 17 | Official leaderboard | |
| NatureBench | 6.67% | #14 of 18 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-Bench Pro | 59.1% | #2 of 24 | Official leaderboard | |
| IOI (Vals) | 67.8% | #6 of 55 | Official leaderboard | |
| WeirdML | 77.7% | #14 of 112 | Official leaderboard | |
| METR Time Horizon | 5.7 h | #6 of 39 | Epoch AI | |
| ALE-Bench | 1,607 | #19 of 121 | Epoch AI | |
| AlgoTune | 1.85× | #3 of 18 | Official leaderboard | |
| OpenHands Index | 64.3% | #6 of 33 | Official leaderboard | |
| SlopCodeBench | 23.5% | #3 of 14 | Official leaderboard | |
| LiveSQLBench | 31% | #8 of 36 | Official leaderboard | |
| METR Time Horizon 1.1 | 5.7 h | #6 of 26 | Official leaderboard | |
| SWE-bench Verified | 76.9% | #8 of 32 | Epoch AI | |
| PostTrainBench | 28.2% | #6 of 23 | Official leaderboard | |
| GSO-Bench | 31.4% | #10 of 30 | Official leaderboard | |
| SWE Atlas Test Writing | 44.4% | #8 of 22 | Official leaderboard | |
| AndroidBench | 74.1% | #12 of 32 | Official leaderboard | |
| Vibe Code Bench | 67.4% | #48 of 115 | Official leaderboard | |
| MLE-bench | 23.3% | #8 of 18 | Model card | |
| SWE Atlas | 40.8% | #10 of 22 | Official leaderboard | |
| SWE Atlas Refactoring | 44.3% | #8 of 17 | Official leaderboard | |
| Code Migration | 35% | #37 of 78 | Official leaderboard | |
| GBAEval | 45.1% | #10 of 21 | Official leaderboard | |
| WebDev Arena | 1,463 | #52 of 105 | Epoch AI | |
| Code Arena | 1,463 | #63 of 124 | Official leaderboard | |
| OpenAI-Proof Q&A | 4.16% | #5 of 9 | Model card | |
| Monorepo-Bench | 59.3% | #3 of 5 | Model card | |
| TasteVal | 21.3× | #13 of 20 | Official leaderboard | |
| ProgramBench | 0% | #17 of 25 | Official leaderboard | |
| DeepSWE | 51.8% | #22 of 28 | Official leaderboard | |
| Image-to-WebDev Arena | 1,436 | #47 of 55 | Official leaderboard | |
| Senior SWE-Bench | 16.8% | #20 of 23 | Official leaderboard | |
| MirrorCode | 15.6% | #25 of 28 | Epoch AI | |
| Vals RSI Index | 0.15 | #23 of 25 | Official leaderboard | |
| PostTrainBench 1.2 | 17.8% | #17 of 17 | Official leaderboard | |
| PostTrainBench 1.1 | 19% | #11 of 11 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| tau2-bench | 98.9% | #1 of 69 | Model card | |
| Terminal-Bench 2.0 | 81.8% | #2 of 43 | Epoch AI | |
| Pencil Puzzle Bench | 70.2% | #3 of 53 | Official leaderboard | |
| Toolathlon | 54.6% | #3 of 50 | Official leaderboard | |
| LLM Chess | 1,322 | #10 of 124 | Official leaderboard | |
| GAIA | 84.7% | #6 of 57 | Official leaderboard | |
| Game Arena Werewolf | 0 | #1 of 8 | Official leaderboard | |
| Game Arena Dark Hex | 379 | #5 of 38 | Official leaderboard | |
| Game Arena Five in a Row | 878 | #3 of 20 | Official leaderboard | |
| Claw-Eval-Live | 63.8% | #2 of 12 | Official leaderboard | |
| LisanBench | 5334.7 | #23 of 133 | Official leaderboard | |
| Game Arena Checkers | 1,094 | #7 of 38 | Official leaderboard | |
| DeepSearchQA | 63.7% | #2 of 10 | Official leaderboard | |
| Claw-Eval | 78.4% | #5 of 23 | Official leaderboard | |
| Game Arena Reversi | 753 | #8 of 36 | Official leaderboard | |
| Chess Puzzles | 48% | #19 of 83 | Epoch AI | |
| SnakeBench | 29 | #42 of 183 | Official leaderboard | |
| Mystery Game Puzzles | 37% | #19 of 80 | Epoch AI | |
| Game Arena Clobber | 452 | #9 of 36 | Official leaderboard | |
| CUA-World-Long | 27.5% | #1 of 4 | Official leaderboard | |
| UniClawBench | 40.8% | #3 of 10 | Official leaderboard | |
| Game Arena Lines of Action | 609 | #11 of 36 | Official leaderboard | |
| Kaggle Game Arena | 352 | #12 of 39 | Official leaderboard | |
| Game Arena Coin Game | 417 | #11 of 35 | Official leaderboard | |
| BrowseComp | 82.7% | #25 of 77 | Model card | |
| Game Arena Dots and Boxes | 950 | #13 of 37 | Official leaderboard | |
| Game Arena Four in a Row | 435 | #14 of 38 | Official leaderboard | |
| Game Arena Chess | 896 | #15 of 39 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 369 | #14 of 36 | Official leaderboard | |
| Game Arena Word Association | 278 | #14 of 36 | Official leaderboard | |
| Claw-Eval Pass^3 | 60.3% | #9 of 23 | Official leaderboard | |
| τ³-bench Banking | 39.4% | #10 of 25 | Official leaderboard | |
| PinchBench 2.0 | 75.7% | #23 of 57 | Official leaderboard | |
| Game Arena Poker | 1.50 | #13 of 31 | Official leaderboard | |
| OSWorld-Verified | 75% | #16 of 37 | Model card | |
| Game Arena Word Art | 603 | #16 of 35 | Official leaderboard | |
| Game Arena Bargaining | 192 | #18 of 36 | Official leaderboard | |
| Game Arena Nine Men's Morris | 607 | #18 of 36 | Official leaderboard | |
| LMArena Search Arena | 1,196 | #18 of 35 | Official leaderboard | |
| Game Arena Go | 854 | #20 of 36 | Official leaderboard | |
| EBR-bench | 25.4% | #16 of 28 | Epoch AI | |
| MCP Atlas | 70.6% | #20 of 34 | Official leaderboard | |
| EdgeBench | 39.3 | #3 of 5 | Official leaderboard | |
| ClawBench | 25.5% | #4 of 6 | Official leaderboard | |
| ApprenticeBench | 11% | #14 of 18 | Official leaderboard | |
| PinchBench | 59.9% | #38 of 48 | Official leaderboard | |
| Long-Horizon-Terminal-Bench (LHTB) | 27.2% | #21 of 25 | Official leaderboard | |
| ARC-AGI-3 | 0.21% | #21 of 24 | Official leaderboard | |
| DeepResearch Bench | 35.1% | #24 of 24 | Epoch AI | |
| iOSWorld | 39.8% | #3 of 3 | Official leaderboard | |
| STATE-Bench | 58.6% | #2 of 2 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HealthBench Consensus | 96.6% | #1 of 17 | Model card | |
| TaxCalcBench | 62.8% | #1 of 17 | Official leaderboard | |
| LegalBench | 86% | #17 of 138 | Official leaderboard | |
| ProfBench | 56.8% | #12 of 72 | Official leaderboard | |
| prinzbench | 69.0 | #5 of 26 | Official leaderboard | |
| MortgageTax | 68.3% | #17 of 86 | Official leaderboard | |
| Vals Finance Agent | 57.1% | #11 of 49 | Official leaderboard | |
| HealthBench | 62.6% | #7 of 30 | Model card | |
| CorpFin v2 | 65.3% | #33 of 119 | Official leaderboard | |
| GDPval | 83% | #3 of 10 | Model card | |
| CaseLaw v2 | 63.8% | #16 of 53 | Official leaderboard | |
| TaxEval v2 | 74% | #43 of 129 | Official leaderboard | |
| HealthBench Hard | 40.1% | #7 of 21 | Model card | |
| Vending-Bench 2 | $6,144 | #24 of 66 | Official leaderboard | |
| PhysicianBench | 27.7% | #5 of 13 | Official leaderboard | |
| APEX-Agents | 52.4% | #27 of 56 | Epoch AI | |
| PRBench | 44.4% | #23 of 44 | Official leaderboard | |
| MedCode | 41.3% | #52 of 96 | Official leaderboard | |
| FORTE | 68.8% | #12 of 21 | Official leaderboard | |
| MedScribe | 77.5% | #63 of 98 | Official leaderboard | |
| Vending-Bench Arena | $2,158 | #25 of 32 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 0% | #69 of 80 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH-Vision (MathVision) | 96.1% | #1 of 105 | Official leaderboard | |
| Math Kangaroo 2025 (MathArena) | 92.5% | #2 of 22 | Official leaderboard | |
| ZeroBench | 23% | #7 of 66 | Official leaderboard | |
| MMMU-Pro | 81.2% | #9 of 72 | Model card | |
| Vision Arena | 1,284 | #19 of 136 | Official leaderboard | |
| VISTA | 50.9% | #8 of 54 | Official leaderboard | |
| Claw-Eval-MM | 54.4% | #2 of 11 | Official leaderboard | |
| EnigmaEval | 16% | #9 of 41 | Epoch AI | |
| ArtifactArena | 1,179 | #5 of 21 | Official leaderboard | |
| ClockBench | 50.6% | #10 of 37 | Official leaderboard | |
| BabyVision | 49.7% | #14 of 27 | Official leaderboard | |
| Furniture Assembly Benchmark | 37.5% | #20 of 35 | Epoch AI | |
| ParseBench | 62.2% | #25 of 42 | Official leaderboard | |
| Blueprint-Bench 2 | 27.1% | #23 of 35 | Official leaderboard | |
| VoxelBench | 1,396 | #44 of 56 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Capture the Flag (Professional) | 88.2% | #3 of 21 | Model card | |
| SMDD-Bench | 40.2% | #1 of 7 | Official leaderboard | |
| ProtocolQA Open-Ended | 42.5% | #4 of 25 | Model card | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 65.3% | #12 of 71 | Official leaderboard | |
| KORA | 58.8% | #12 of 42 | Official leaderboard | |
| SEC-bench Pro | 31.7% | #2 of 6 | Official leaderboard | |
| SEC-bench Pro 260617 | 39% | #2 of 6 | Official leaderboard | |
| CyberGym | 79% | #12 of 33 | Official leaderboard | |
| Destructive action avoidance | 86% | #3 of 8 | Model card | |
| CyberBench (Vals) | 72.9% | #10 of 26 | Official leaderboard | |
| CyberGym-E2E | 65.9% | #2 of 5 | Official leaderboard | |
| TroubleshootingBench | 35.8% | #9 of 19 | Model card | |
| Dictatorship Eval | 75.7% | #10 of 20 | Official leaderboard | |
| CVE-Bench | 86.3% | #7 of 14 | Model card | |
| Cyber Range | 73.3% | #5 of 10 | Model card | |
| First-Person Fairness evaluation | 0.88% | #5 of 10 | Model card | |
| CyScenarioBench | 11% | #8 of 14 | Model card | |
| Agent JSK | 99.8% | #4 of 7 | Model card | |
| Multimodal Troubleshooting Virology | 50.4% | #16 of 26 | Model card | |
| BinaryAudit | 17.5% | #14 of 20 | Official leaderboard | |
| ExploitGym | 7.02% | #13 of 18 | Official leaderboard | |
| BiomniBench | 55.2 | #7 of 8 | Official leaderboard | |
| CoT-Control | 0.30% | #7 of 8 | Model card | |
| Tacit Knowledge and Troubleshooting | 65% | #26 of 29 | Model card | |
| PCT (Polarized Contrastive Pairs) | 59.4% | #42 of 46 | Official leaderboard | |
| PlugInject | 97.8% | #2 of 2 | Model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ACCoRD | 80.0 | #12 of 150 | Official leaderboard | |
| Conceptual Reasoning Index | 65.3 | #20 of 149 | Official leaderboard | |
| LMCA | 52% | #23 of 139 | Epoch AI | |
| CAIS Text Capabilities Index | 41.3 | #16 of 66 | Official leaderboard | |
| CAIS Vision Capabilities Index | 57.7 | #19 of 53 | Official leaderboard | |
| CAIS Risk Index | 49.2 | #28 of 65 | Official leaderboard | |
| SAGE (Vals) | 43.3% | #49 of 81 | Official leaderboard |
OpenAI's models over time
OpenAI's models in color, against every other model on the Epoch Capabilities Index, by release date.
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 31 of them aren't in the launch post.
GPT-5.4 Thinking System Card 5 Mar 2026
Health
- HealthBench 62.6
- HealthBench Hard 40.1
- HealthBench Consensus 96.6
Safety
- OpenAI Production Benchmarks and safe-completion evals
- OpenAI Dynamic Mental Health Benchmarks
- Disallowed Content Evaluations with Representative Prompts
- Image input safety evaluations
- Destructive action avoidance 86
- Destructive action reversion on long rollouts
- User Confirmations During Computer Use
- OpenAI cyber safety evaluations (incl. Daybreak Blue)
Robustness
Bias
Alignment
Monitorability
Bio and chem risk
Cyber risk
- Capture the Flag (Professional) 88.2
- CVE-Bench 86.3
- Cyber Range 73.3
- Atomic Challenges
- CyScenarioBench 11
AI R&D risk
- Monorepo-Bench 59.3
- MLE-bench 23.3
- OpenAI-Proof Q&A 4.16