# GPT-4o vs gpt-oss-20b

> gpt-oss-20b is the stronger model overall, scoring 32.1 to 28.5 on the Noometry Index.

- Canonical page: https://noometry.com/compare/gpt-4o-vs-gpt-oss-20b
- Last updated: 2026-10-10
- Shared benchmarks: 29

## Summary

- They share 29 benchmarks with published results for both. GPT-4o scores higher in 5 categories and gpt-oss-20b in 4 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where gpt-oss-20b leads 37.6 to 10.2.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 6.4% for GPT-4o and 65.3% for gpt-oss-20b.
- gpt-oss-20b is cheaper at $0.018 / $0.09 per million input/output tokens, against $2.50 / $10 for GPT-4o.
- gpt-oss-20b accepts more context: 131K tokens versus 128K.
- gpt-oss-20b has downloadable open weights; the other is API-only.

## Snapshot

| | GPT-4o | gpt-oss-20b |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 28.5 | 32.1 |
| Rank | 316 | 249 |
| Context | 128K | 131K |
| Input $/M | $2.50 | $0.018 |
| Output $/M | $10 | $0.09 |
| Weights | Proprietary | Open |

## Coding

- GPT-4o: 25.1 (#323)
- gpt-oss-20b: 37.1 (#185)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| WeirdML | 25.1% | 40.9% |
| LMArena Coding | 1297 | 1307 |
| SWE-bench Verified | 31% | — |
| SWE-bench Verified (bash only) | 21.6% | — |
| Aider Polyglot | 45.3% | — |
| SciCode | — | 34.4% |
| GSO | 0% | — |
| BigCodeBench Instruct | 51.1% | — |
| LiveBench Coding | 51.4% | — |
| BigCodeBench Complete | 61.1% | — |
| CadEval | 26% | — |
| ALE-Bench | — | 566.05 |
| HumanEval+ | 87.2% | — |
| MBPP+ | 72.2% | — |

## Agentic & Tool Use

- GPT-4o: 20.8 (#133)
- gpt-oss-20b: 9.2 (#146)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| Terminal-Bench | — | 3.4% |
| GDPval | 9.9% | — |
| TheAgentCompany | 8.6% | — |
| Cybench | 12.5% | — |
| BALROG | 32.3% | — |
| LMArena Search | 1006 | — |
| METR Time Horizons | 40.8% | — |

## Reasoning

- GPT-4o: 9.4 (#338)
- gpt-oss-20b: 18.8 (#251)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| CritPt | 0% | 1.4% |
| Chess Puzzles | 13% | 4% |
| LMArena Hard Prompts | 1281 | 1274 |
| DTBench | 64.5% | 68% |
| LMCA | 16.6% | 14.5% |
| Epoch Capabilities Index | 128.97 | 137.81 |
| ARC-AGI-2 | 0% | — |
| SimpleBench | 17.8% | — |
| Kagi LLM Benchmark | — | 53.2% |
| ARC-AGI-1 | 4.5% | — |
| EnigmaEval | 0.8% | — |
| LiveBench Reasoning | 55.8% | — |
| LiveBench Data Analysis | 60.9% | — |
| ForecastBench | 57.7 | — |
| LiveBench | 55.3% | — |

## Math

- GPT-4o: 10.2 (#303)
- gpt-oss-20b: 37.6 (#93)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 6.4% | 65.3% |
| Omni-MATH | 29.3% | 56.5% |
| LMArena Math | 1285 | 1317 |
| FrontierMath (Tiers 1-3) | 0.4% | — |
| LiveBench Math | 49.5% | — |
| MATH Level 5 | 53.3% | — |
| FrontierMath (Feb 2025 set) | 0.3% | — |

## Knowledge

- GPT-4o: 28.8 (#235)
- gpt-oss-20b: 34.6 (#186)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| GPQA Diamond | 49.2% | 60.8% |
| MMLU-Pro | 71.3% | 74% |
| GPQA (HELM) | 52% | 59.4% |
| LMArena Expert | 1250 | 1258 |
| Humanity's Last Exam | 2.7% | — |
| SimpleQA Verified | 26% | — |
| Confabulations | 15.3% | — |
| Vectara Hallucination Rate | 9.6% | — |
| MMLU | 88.1% | — |

## Multimodal

- GPT-4o: 33.9 (#85)
- gpt-oss-20b: —

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| LMArena Vision | 1137 | — |
| Video-MME | 71.9% | — |
| GeoBench | 71% | — |
| VPCT | 40% | — |
| ScienceQA | 88.5% | — |

## Multilingual

- GPT-4o: 43.4 (#181)
- gpt-oss-20b: 42.4 (#192)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| LMArena Non-English | 1283 | 1269 |
| LMArena Chinese | 1277 | 1311 |
| LMArena German | 1282 | 1249 |
| LMArena Japanese | 1257 | 1245 |
| LMArena Korean | 1234 | 1239 |
| LMArena Russian | 1286 | 1279 |
| LMArena Spanish | 1292 | 1267 |
| LMArena French | 1303 | — |

## Instruction Following

- GPT-4o: 66.8 (#202)
- gpt-oss-20b: 62.0 (#235)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| IFEval | 81.7% | 73.2% |
| LMArena Instruction Following | 1278 | 1236 |
| LiveBench Instruction Following | 68.6% | — |

## Long Context

- GPT-4o: 39.4 (#174)
- gpt-oss-20b: 37.9 (#203)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| LMArena Longer Query | 1289 | 1251 |
| Fiction.LiveBench | 66.7% | — |

## Writing & Preference

- GPT-4o: 52.8 (#160)
- gpt-oss-20b: 35.8 (#260)

| Benchmark | GPT-4o | gpt-oss-20b |
|---|---|---|
| LMArena Text | 1300 | 1288 |
| LMArena Creative Writing | 1291 | 1201 |
| WildBench | 82.8% | 73.7% |
| LMArena Multi-Turn | 1302 | 1268 |
| Short-Story Creative Writing | 81.8% | — |
| EQ-Bench Creative Writing | — | 666 |
| LiveBench Language | 47.6% | — |

## FAQ

### Is GPT-4o better than gpt-oss-20b?

gpt-oss-20b is the stronger model overall, scoring 32.1 to 28.5 on the Noometry Index.

### Which is cheaper, GPT-4o or gpt-oss-20b?

gpt-oss-20b is cheaper. It lists at $0.018 per million input tokens and $0.09 per million output tokens; GPT-4o lists at $2.50 and $10.

### Is GPT-4o or gpt-oss-20b better for coding?

gpt-oss-20b scores higher on coding benchmarks: 37.1 versus 25.1 in the Noometry coding category.

### Which has the bigger context window?

gpt-oss-20b does, with 131K tokens against 128K.

### How many benchmarks do GPT-4o and gpt-oss-20b share?

29 benchmarks have published results for both models. GPT-4o has 72 scored results on Noometry and gpt-oss-20b has 34.