Model comparison
GPT-5.2 vs GPT-5.4
GPT-5.4 is the stronger model overall, scoring 59.4 to 54.1 on the Noometry Index.
Last verified . 57 shared benchmarks.
Summary
- They share 57 benchmarks with published results for both. GPT-5.2 scores higher in 1 category and GPT-5.4 in 9 categories; 10 gaps are clear of the uncertainty.
- The widest gap is in math, where GPT-5.4 leads 73.5 to 60.0.
- The biggest single-benchmark swing is ProofBench: 15% for GPT-5.2 and 56% for GPT-5.4.
- GPT-5.2 is cheaper at $1.75 / $14 per million input/output tokens, against $2.50 / $15 for GPT-5.4.
- GPT-5.4 accepts more context: 1.05M tokens versus 400K.
Side by side
| GPT-5.2 | GPT-5.4 | |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 54.1 | 59.4 |
| Released | 2025-12-11 | 2026-03-05 |
| Weights | Proprietary | Proprietary |
| Context window | 400K | 1.05M |
| Max output | 128K | 128K |
| Input $ / M tokens | $1.75 | $2.50 |
| Output $ / M tokens | $14 | $15 |
| Results tracked | 67 | 68 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-5.4 leads
GPT-5.2: 51.6 (#37), GPT-5.4: 52.6 (#33)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| SWE-bench Verified | 73.8% | 76.9% |
| LMArena WebDev | 1416 | 1465 |
| GSO | 27.4% | 31.4% |
| WeirdML | 72.2% | 77.7% |
| LMArena Coding | 1447 | 1497 |
| ALE-Bench | 1,294 | 1,607 |
| AlgoTune | 2.05 | 1.85 |
| DeepSWE | — | 51.8% |
| SWE-bench Verified (bash only) | 72.8% | — |
| SWE-bench Multilingual | 66.7% | — |
| SciCode | — | 56.6% |
| MirrorCode | — | 15.6% |
Agentic & Tool Use GPT-5.4 leads
GPT-5.2: 40.2 (#24), GPT-5.4: 46.5 (#13)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| Terminal-Bench | 64.9% | 81.8% |
| τ²-bench Banking | 32.2% | 39.4% |
| DeepResearch Bench | 41.1% | 35.1% |
| LMArena Search | 1207 | 1197 |
| METR Time Horizons | 75.3% | 74.3% |
| Vending-Bench 2 | 3,591 | 6,144 |
| APEX-Agents | — | 52.4% |
| Berkeley Function Calling Leaderboard | 55.9% | — |
| GDPval | 49.7% | — |
| Remote Labor Index | 2.5% | — |
| τ²-bench Airline | 83% | — |
| τ²-bench Retail | 81.6% | — |
| τ²-bench Telecom | 89.7% | — |
| PostTrainBench | — | 19% |
| GBAEval | — | 45.1% |
Reasoning GPT-5.4 leads
GPT-5.2: 50.2 (#35), GPT-5.4: 61.8 (#19)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| ARC-AGI-2 | 52.9% | 74% |
| Kagi LLM Benchmark | 73.3% | 63.8% |
| NYT Connections (extended) | 83.6% | 91.3% |
| ARC-AGI-1 | 86.2% | 93.7% |
| Chess Puzzles | 49% | 44% |
| EnigmaEval | 10.4% | 16% |
| EBR-Bench | 23% | 25.4% |
| LMArena Hard Prompts | 1445 | 1485 |
| Mystery Game Puzzles | 23% | 37% |
| DTBench | 90.9% | 94.4% |
| LMCA | 43.9% | 52% |
| Epoch Capabilities Index | 153.45 | 156.81 |
| ForecastBench | 60.1 | 59.5 |
| SimpleBench | 45.8% | — |
| CritPt | — | 23.4% |
| Thematic Generalization | — | 80% |
Math GPT-5.4 leads
GPT-5.2: 60.0 (#38), GPT-5.4: 73.5 (#19)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| FrontierMath (Tiers 1-3) | 67.4% | 78.6% |
| FrontierMath Tier 4 | 31.7% | 49% |
| MathArena Final-Answer Competitions | 72% | 83.1% |
| OTIS Mock AIME 2024-2025 | 96.1% | 97.8% |
| ProofBench | 15% | 56% |
| LMArena Math | 1440 | 1488 |
| FrontierMath (Feb 2025 set) | 40.7% | 47.6% |
| FrontierMath Tier 4 (v1) | 18.8% | 27.1% |
Knowledge GPT-5.4 leads
GPT-5.2: 59.3 (#32), GPT-5.4: 65.3 (#14)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| GPQA Diamond | 91.4% | 93.3% |
| Humanity's Last Exam | 27.8% | 36.2% |
| SimpleQA Verified | 37.1% | 45.1% |
| Vectara Hallucination Rate | 8.4% | 7% |
| LMArena Expert | 1445 | 1507 |
Multimodal GPT-5.2 leads
GPT-5.2: 51.3 (#7), GPT-5.4: 43.7 (#20)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| LMArena Vision | 1268 | 1303 |
| Furniture Assembly | 38.3% | 37.5% |
| LMArena Document | 1405 | 1471 |
| VPCT | 84% | — |
| Blueprint-Bench 2 | — | 27.1% |
Multilingual GPT-5.4 leads
GPT-5.2: 53.4 (#67), GPT-5.4: 56.2 (#23)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| LMArena Non-English | 1425 | 1465 |
| LMArena Chinese | 1460 | 1519 |
| LMArena French | 1455 | 1493 |
| LMArena German | 1448 | 1472 |
| LMArena Japanese | 1420 | 1485 |
| LMArena Korean | 1392 | 1448 |
| LMArena Russian | 1440 | 1480 |
| LMArena Spanish | 1433 | 1454 |
Instruction Following GPT-5.4 leads
GPT-5.2: 74.7 (#89), GPT-5.4: 77.1 (#27)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| LMArena Instruction Following | 1417 | 1469 |
Long Context GPT-5.4 leads
GPT-5.2: 44.0 (#78), GPT-5.4: 50.3 (#8)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| CL-bench | 18.2% | 27.9% |
| LMArena Longer Query | 1428 | 1473 |
| CL-bench Life | — | 21.7% |
Writing & Preference GPT-5.4 leads
GPT-5.2: 66.8 (#32), GPT-5.4: 71.9 (#17)
| Benchmark | GPT-5.2 | GPT-5.4 |
|---|---|---|
| LMArena Text | 1439 | 1469 |
| LMArena Creative Writing | 1401 | 1439 |
| EQ-Bench Creative Writing | 1703 | 1840 |
| LMArena Multi-Turn | 1458 | 1482 |
| EQ-Bench 4 | — | 1272 |
Frequently asked questions
Is GPT-5.2 better than GPT-5.4?
GPT-5.4 is the stronger model overall, scoring 59.4 to 54.1 on the Noometry Index.
Which is cheaper, GPT-5.2 or GPT-5.4?
GPT-5.2 is cheaper. It lists at $1.75 per million input tokens and $14 per million output tokens; GPT-5.4 lists at $2.50 and $15.
Is GPT-5.2 or GPT-5.4 better for coding?
GPT-5.4 scores higher on coding benchmarks: 52.6 versus 51.6 in the Noometry coding category.
Which has the bigger context window?
GPT-5.4 does, with 1.05M tokens against 400K.
How many benchmarks do GPT-5.2 and GPT-5.4 share?
57 benchmarks have published results for both models. GPT-5.2 has 67 scored results on Noometry and GPT-5.4 has 68.