Up to four models
GPT-5.4 vs GPT-5.2
On the index (GPT-4 = 100), GPT-5.4 scores 166 and GPT-5.2 162. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
GPT-5.4
166150–182
WeightsClosed
ReleasedMar 2026
Price in / out$2.5 / $15
Context1.05M
Core tests taken7 of 9
GPT-5.2
162146–178
WeightsClosed
ReleasedDec 2025
Price in / out$1.75 / $14
Context400K
Core tests taken6 of 9
GPT-5.4 and GPT-5.2 are too close to call (4 points apart, within the uncertainty), ahead on 5 of 6 shared core tests, at a similar price.
Test by testthird-party results
GPT-5.4GPT-5.2
Every result as a table
| Test | GPT-5.4 | GPT-5.2 |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 93.3% | 91.4% |
| Humanity's Last Exam | 36.2% | 27.8% |
| GPQA Diamond (Vals AI) | 91.7% | 91.7% |
| MMLU Pro | 87.5% | 86.2% |
| MMMU Pro | 87.5% | 86.7% |
| CL-bench | 27.9% | 18.2% |
| EBR-bench | 25.4% | 23.0% |
| SimpleQA Verified | 45.1% | 37.1% |
| Maths | ||
| FrontierMath (tiers 1-3) | 78.6% | 67.4% |
| AIME-style maths (OTIS mock) | 97.8% | 96.1% |
| Chess puzzles | 48.0% | 48.0% |
| FrontierMath (2025 set) | 47.6% | 40.7% |
| FrontierMath tier 4 (2025 set) | 27.1% | 18.8% |
| FrontierMath (tier 4) | 49.0% | 31.7% |
| Coding | ||
| SWE-bench Verified | 76.9% | 73.8% |
| SWE-bench Pro (public set) | 59.1% | 29.9% |
| LiveCodeBench | 84.1% | 85.4% |
| SWE-bench | 78.2% | 75.8% |
| Vibe Code Bench v1.1 | 67.4% | 53.5% |
| GSO-Bench | 31.4% | 27.5% |
| MirrorCode | 23.3% | 32.2% |
| WeirdML | 77.7% | 72.2% |
| Agents | ||
| DTBench | 94.4% | 90.9% |
| DeepResearch Bench | 35.1% | 41.1% |
| Furniture assembly | 37.5% | 38.3% |
| LMCA | 52.0% | 43.9% |
| Mystery game puzzles | 37.0% | 23.0% |
| Professional work | ||
| CorpFin v2 | 65.3% | 65.9% |
| LegalBench | 86.0% | 82.8% |
| MedCode | 41.3% | 49.8% |
| MedScribe | 77.5% | 84.4% |
| MortgageTax | 68.3% | 67.1% |
| TaxEval v2 | 74.0% | 75.8% |
| Novel problems | ||
| ARC-AGI-2 | 74.0% | 72.9% |
| ARC-AGI-1 | 93.7% | 94.5% |
Task length · METR45 sec45 sec
Tests at least two of these models have taken. Bold rows are the core tests.