Model comparison
GPT-3.5-turbo vs GPT-4
GPT-4 is the stronger model overall, scoring 29.1 to 23.2 on the Noometry Index. GPT-3.5-turbo costs 50× less per token, which makes it the better buy when GPT-4's lead doesn't matter for your workload.
Last verified . 37 shared benchmarks.
Summary
- They share 37 benchmarks with published results for both. GPT-3.5-turbo scores higher in 0 categories and GPT-4 in 8 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in writing & preference, where GPT-4 leads 34.9 to 25.3.
- The biggest single-benchmark swing is DTBench: 48.5% for GPT-3.5-turbo and 62.7% for GPT-4.
- GPT-3.5-turbo is cheaper at $0.50 / $1.50 per million input/output tokens, against $30 / $60 for GPT-4.
- GPT-3.5-turbo accepts more context: 16K tokens versus 8K.
Side by side
| GPT-3.5-turbo | GPT-4 | |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 23.2 | 29.1 |
| Released | 2023-03-01 | 2023-03-14 |
| Weights | Proprietary | Proprietary |
| Context window | 16K | 8K |
| Max output | 4K | 8K |
| Input $ / M tokens | $0.50 | $30 |
| Output $ / M tokens | $1.50 | $60 |
| Results tracked | 44 | 38 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-4 leads
GPT-3.5-turbo: 23.9 (#331), GPT-4: 31.6 (#283)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| WeirdML | 3.5% | 12.4% |
| BigCodeBench Instruct | 39.1% | 46% |
| LMArena Coding | 1136 | 1254 |
| BigCodeBench Complete | 50.6% | 57.2% |
| HumanEval+ | 70.7% | 79.3% |
| MBPP+ | 69.7% | — |
Agentic & Tool Use Not comparable
GPT-3.5-turbo: —, GPT-4: —
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| METR Time Horizons | 21.5% | 36.1% |
Reasoning GPT-4 leads
GPT-3.5-turbo: 13.8 (#332), GPT-4: 17.8 (#289)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| Chess Puzzles | 0% | 4% |
| LMArena Hard Prompts | 1108 | 1241 |
| Mystery Game Puzzles | 3% | 12% |
| DTBench | 48.5% | 62.7% |
| LMCA | 9.7% | 17.1% |
| BIG-Bench Hard | 61.6% | 75.1% |
| Epoch Capabilities Index | 118.55 | 125.89 |
| ForecastBench | 50.4 | 57.8 |
| WinoGrande | 81.6% | 87.5% |
| Adversarial NLI | 58.1% | — |
| CommonsenseQA 2.0 | 57% | — |
| HellaSwag | — | 95.3% |
Math GPT-4 leads
GPT-3.5-turbo: 6.3 (#327), GPT-4: 10.8 (#309)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 2.2% | 1.1% |
| LMArena Math | 1142 | 1269 |
| MATH Level 5 | 15.9% | 23% |
| GSM8K | 57.8% | 92% |
| FrontierMath (Tiers 1-3) | 0% | — |
Knowledge GPT-4 leads
GPT-3.5-turbo: 10.0 (#303), GPT-4: 18.4 (#282)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| GPQA Diamond | 28% | 35.7% |
| LMArena Expert | 1070 | 1211 |
| MMLU | 71.4% | 86.4% |
| TriviaQA | 85.8% | 84.8% |
| ARC (AI2) Challenge | 87.4% | — |
| BoolQ | 87% | — |
| OpenBookQA | 86% | — |
Multilingual GPT-4 leads
GPT-3.5-turbo: 31.5 (#258), GPT-4: 40.6 (#215)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| LMArena Non-English | 1108 | 1246 |
| LMArena Chinese | 1075 | 1242 |
| LMArena French | 1118 | 1283 |
| LMArena German | 1090 | 1251 |
| LMArena Japanese | 1043 | 1209 |
| LMArena Korean | 1019 | 1184 |
| LMArena Russian | 1123 | 1251 |
| LMArena Spanish | 1121 | 1261 |
Instruction Following GPT-4 leads
GPT-3.5-turbo: 57.9 (#262), GPT-4: 65.3 (#222)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| LMArena Instruction Following | 1119 | 1241 |
Long Context GPT-4 leads
GPT-3.5-turbo: 34.0 (#254), GPT-4: 37.7 (#212)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| LMArena Longer Query | 1121 | 1244 |
Writing & Preference GPT-4 leads
GPT-3.5-turbo: 25.3 (#305), GPT-4: 34.9 (#268)
| Benchmark | GPT-3.5-turbo | GPT-4 |
|---|---|---|
| LMArena Text | 1125 | 1263 |
| LMArena Creative Writing | 1092 | 1244 |
| EQ-Bench Creative Writing | 451 | 752 |
| LMArena Multi-Turn | 1117 | 1257 |
Frequently asked questions
Is GPT-3.5-turbo better than GPT-4?
GPT-4 is the stronger model overall, scoring 29.1 to 23.2 on the Noometry Index. GPT-3.5-turbo costs 50× less per token, which makes it the better buy when GPT-4's lead doesn't matter for your workload.
Which is cheaper, GPT-3.5-turbo or GPT-4?
GPT-3.5-turbo is cheaper. It lists at $0.50 per million input tokens and $1.50 per million output tokens; GPT-4 lists at $30 and $60.
Is GPT-3.5-turbo or GPT-4 better for coding?
GPT-4 scores higher on coding benchmarks: 31.6 versus 23.9 in the Noometry coding category.
Which has the bigger context window?
GPT-3.5-turbo does, with 16K tokens against 8K.
How many benchmarks do GPT-3.5-turbo and GPT-4 share?
37 benchmarks have published results for both models. GPT-3.5-turbo has 44 scored results on Noometry and GPT-4 has 38.