Introducing GPT-5.4 mini and nano 17 Mar 2026
- SWE-Bench Pro (Public) 54.4
- Terminal-Bench 2.0 60
- Toolathlon 42.9
- GPQA Diamond 88
- OSWorld-Verified 72.1
GPT-5.4 Mini
GPT-5.4 Mini places in the top three on OpenAI-Proof Q&A.
107
Benchmarks with a score
0
Ranked first, out of five or more models
1
In the top three
OpenAI Mar 2026
GPT-5.4 Mini
Closed weights Small
148.8
#69 of 271
Epoch Capabilities Index
- Toolathlon 42.9% #16 /50
- ZeroBench 11% #28 /66
- Harvey LAB (Legal Agent Benchmark) 0% #70 /80
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 | |
|---|---|---|---|---|
| Wer wird Millionär? LLM Benchmark | 225000.0 | #11 of 76 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 5.50% | #20 of 104 | Official leaderboard | |
| LMArena (Elo) | 1,447 | #67 of 333 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 73.0 | #33 of 119 | Official leaderboard | |
| HELM Arabic Enterprise | 70.2 | #7 of 23 | Official leaderboard | |
| Judgemark v2.1 | 81.1 | #13 of 42 | Official leaderboard | |
| Creative Writing v3 | 72.5 | #39 of 120 | Official leaderboard | |
| SEA-HELM | 72.8 | #23 of 62 | Official leaderboard | |
| HELM Arabic | 79.0 | #13 of 28 | Official leaderboard | |
| BullshitBench V2 | 32% | #68 of 122 | Official leaderboard | |
| BullshitBench | 25.4% | #60 of 107 | Official leaderboard | |
| Extended NYT Connections | 61.8% | #58 of 97 | Official leaderboard | |
| SimpleQA Verified | 29.4% | #66 of 84 | Epoch AI | |
| Judgemark v4 | 27.3 | #38 of 40 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FrontierMath | 28.3% | #20 of 70 | Epoch AI | |
| GPQA Diamond | 86.9% | #64 of 219 | Epoch AI | |
| ForecastBench | 61.0 | #26 of 86 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 88.9% | #64 of 202 | Epoch AI | |
| CritPt | 10% | #58 of 160 | Epoch AI | |
| ProofBench | 21% | #29 of 59 | Official leaderboard | |
| DTBench | 66.7% | #92 of 178 | Official leaderboard | |
| ARC-AGI-2 | 18.9% | #51 of 95 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 51.2% | #54 of 89 | Epoch AI | |
| ARC-AGI | 63.7% | #57 of 92 | Official leaderboard | |
| BixBench3 | 36.4 | #15 of 20 | Official leaderboard | |
| FrontierMath Tier 4 | 2.08% | #45 of 55 | Epoch AI | |
| FrontierMath Tier 4 v2 | 9.76% | #60 of 71 | Epoch AI |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WeirdML | 60.3% | #34 of 112 | Official leaderboard | |
| FrontierCode | 17.8% | #4 of 12 | Official leaderboard | |
| OpenAI-Proof Q&A | 7.50% | #3 of 9 | Model card | |
| ALE-Bench | 1,189 | #43 of 121 | Epoch AI | |
| Vibe Code Bench | 48% | #62 of 115 | Official leaderboard | |
| SWE-Bench Pro | 54.4% | #49 of 77 | Model card | |
| WebDev Arena | 1,397 | #68 of 105 | Epoch AI | |
| Code Arena | 1,397 | #84 of 124 | Official leaderboard | |
| IOI (Vals) | 6.42% | #39 of 55 | Official leaderboard | |
| SlopCodeBench | 13.8% | #10 of 14 | Official leaderboard | |
| ProgramBench | 0% | #18 of 25 | Official leaderboard | |
| FrontierCode 1.1 | 27% | #6 of 8 | Official leaderboard | |
| Code Migration | 12.9% | #61 of 78 | Official leaderboard | |
| SWE-sweep | 0.50% | #9 of 10 | Official leaderboard | |
| Monorepo-Bench | 54% | #5 of 5 | Model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 60% | #19 of 61 | Model card | |
| Toolathlon | 42.9% | #16 of 50 | Official leaderboard | |
| PinchBench 2.0 | 79.2% | #19 of 57 | Official leaderboard | |
| LLM Chess | 416 | #46 of 124 | Official leaderboard | |
| PinchBench | 71.4% | #18 of 48 | Official leaderboard | |
| LisanBench | 1140.7 | #54 of 133 | Official leaderboard | |
| GAIA | 64.5% | #24 of 57 | Official leaderboard | |
| Game Arena Five in a Row | 497 | #10 of 20 | Official leaderboard | |
| Game Arena Poker | -0.17 | #16 of 31 | Official leaderboard | |
| OSWorld-Verified | 72.1% | #22 of 37 | Model card | |
| Chess Puzzles | 23% | #54 of 83 | Epoch AI | |
| Game Arena Four in a Row | 206 | #28 of 38 | Official leaderboard | |
| Game Arena Clobber | 145 | #27 of 36 | Official leaderboard | |
| Game Arena Lines of Action | 392 | #27 of 36 | Official leaderboard | |
| Game Arena Werewolf | -6 | #6 of 8 | Official leaderboard | |
| MyPCBench | 23.9% | #7 of 9 | Official leaderboard | |
| UniClawBench | 21.5% | #8 of 10 | Official leaderboard | |
| SkillsBench | 41.4% | #17 of 21 | Official leaderboard | |
| Game Arena Dots and Boxes | 612 | #30 of 37 | Official leaderboard | |
| Mystery Game Puzzles | 11% | #65 of 80 | Epoch AI | |
| Game Arena Word Association | 64 | #30 of 36 | Official leaderboard | |
| ClawBench | 24.8% | #5 of 6 | Official leaderboard | |
| Game Arena Dark Hex | 98 | #33 of 38 | Official leaderboard | |
| Game Arena Chess | 435 | #34 of 39 | Official leaderboard | |
| Game Arena Word Art | 126 | #31 of 35 | Official leaderboard | |
| MCP Atlas | 56.7% | #31 of 34 | Official leaderboard | |
| SnakeBench | 15 | #167 of 183 | Official leaderboard | |
| Game Arena Nine Men's Morris | 208 | #33 of 36 | Official leaderboard | |
| Game Arena Reversi | 95 | #33 of 36 | Official leaderboard | |
| DeepResearch Bench | 36.3% | #22 of 24 | Epoch AI | |
| Game Arena Checkers | 279 | #35 of 38 | Official leaderboard | |
| Kaggle Game Arena | 140 | #37 of 39 | Official leaderboard | |
| Game Arena Coin Game | 73 | #34 of 35 | Official leaderboard | |
| Game Arena Go | 121 | #35 of 36 | Official leaderboard | |
| Game Arena Bargaining | 0 | #36 of 36 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 0 | #36 of 36 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vals Finance Agent | 53.4% | #20 of 49 | Official leaderboard | |
| CorpFin v2 | 60.9% | #63 of 119 | Official leaderboard | |
| MortgageTax | 63.5% | #49 of 86 | Official leaderboard | |
| TaxEval v2 | 71.2% | #75 of 129 | Official leaderboard | |
| Vals Finance Agent v2 | 45.4% | #55 of 80 | Official leaderboard | |
| Excel Modeling Benchmark | 45.4% | #58 of 76 | Official leaderboard | |
| Tax Agent Bench | 8.81% | #66 of 79 | Official leaderboard | |
| Legal Research Bench | 12.5% | #76 of 88 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 0% | #70 of 80 | Official leaderboard | |
| CaseLaw v2 | 51.7% | #47 of 53 | Official leaderboard | |
| PolicyBench | 73.3% | #47 of 50 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,250 | #45 of 136 | Official leaderboard | |
| ZeroBench | 11% | #28 of 66 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Capture the Flag (Professional) | 81.3% | #7 of 21 | Model card | |
| Tacit Knowledge and Troubleshooting | 71.7% | #16 of 29 | Model card | |
| CVE-Bench | 83.3% | #8 of 14 | Model card | |
| ProtocolQA Open-Ended | 33.6% | #15 of 25 | Model card | |
| TroubleshootingBench | 31.9% | #14 of 19 | Model card | |
| CRSBench | 54.9% | #4 of 5 | Official leaderboard | |
| Multimodal Troubleshooting Virology | 46.5% | #22 of 26 | Model card | |
| PCT (Polarized Contrastive Pairs) | 61.1% | #44 of 46 | Official leaderboard |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SAGE (Vals) | 50.8% | #15 of 81 | Official leaderboard | |
| ACCoRD | 70.5 | #49 of 150 | Official leaderboard | |
| LMCA | 40.8% | #50 of 139 | Epoch AI | |
| Conceptual Reasoning Index | 51.9 | #61 of 149 | Official leaderboard | |
| CAIS Text Capabilities Index | 26.6 | #38 of 66 | Official leaderboard | |
| CAIS Risk Index | 53.9 | #38 of 65 | Official leaderboard | |
| Vals Multimodal Index | 54.2% | #21 of 33 | Official leaderboard | |
| CAIS Vision Capabilities Index | 46.6 | #34 of 53 | 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. 11 of them aren't in the launch post.
Safety
Monitorability
Bio and chem risk
Cyber risk
- Capture the Flag (Professional) 81.3
- CVE-Bench 83.3
AI R&D risk
- Monorepo-Bench 54
- OpenAI-Proof Q&A 7.5