# gpt-oss-120b vs gpt-oss-20b

> gpt-oss-120b is the stronger model overall, scoring 36.3 to 32.5 on the Noometry Index. gpt-oss-20b costs 2.0× less per token, which makes it the better buy when gpt-oss-120b's lead doesn't matter for your workload.

- Canonical page: https://noometry.com/compare/gpt-oss-120b-vs-gpt-oss-20b
- Last updated: 2026-10-12
- Shared benchmarks: 34

## Summary

- They share 34 benchmarks with published results for both. gpt-oss-120b scores higher in 7 categories and gpt-oss-20b in 2 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where gpt-oss-120b leads 52.5 to 39.4.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 88.9% for gpt-oss-120b and 65.3% for gpt-oss-20b.
- gpt-oss-20b is cheaper at $0.018 / $0.09 per million input/output tokens, against $0.037 / $0.17 for gpt-oss-120b.

## Snapshot

| | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 36.3 | 32.5 |
| Rank | 217 | 255 |
| Context | 131K | 131K |
| Input $/M | $0.037 | $0.018 |
| Output $/M | $0.17 | $0.09 |
| Weights | Open | Open |

## Coding

- gpt-oss-120b: 33.5 (#256)
- gpt-oss-20b: 37.6 (#192)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| SciCode | 36% | 34.4% |
| WeirdML | 48.2% | 40.9% |
| LMArena Coding | 1380 | 1306 |
| ALE-Bench | 575.62 | 566.05 |
| SWE-bench Verified (bash only) | 26% | — |
| Aider Polyglot | 41.8% | — |
| AlgoTune | 1.41 | — |

## Agentic & Tool Use

- gpt-oss-120b: 12.2 (#153)
- gpt-oss-20b: 9.3 (#154)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Terminal-Bench | 18.7% | 3.4% |
| APEX-Agents | 4.4% | — |
| METR Time Horizons | 56.6% | — |
| Vending-Bench 2 | -21.53 | — |

## Reasoning

- gpt-oss-120b: 20.0 (#245)
- gpt-oss-20b: 19.3 (#261)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Kagi LLM Benchmark | 58.6% | 53.2% |
| CritPt | 1.1% | 1.4% |
| Chess Puzzles | 20% | 4% |
| LMArena Hard Prompts | 1364 | 1274 |
| DTBench | 76.3% | 68% |
| LMCA | 22.1% | 14.5% |
| Epoch Capabilities Index | 139.93 | 137.82 |
| SimpleBench | 22.1% | — |
| Mystery Game Puzzles | 2% | — |
| Surface Evolver Bench | 25% | — |

## Math

- gpt-oss-120b: 52.5 (#50)
- gpt-oss-20b: 39.4 (#103)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 88.9% | 65.3% |
| Omni-MATH | 68.8% | 56.5% |
| LMArena Math | 1389 | 1317 |

## Knowledge

- gpt-oss-120b: 42.4 (#96)
- gpt-oss-20b: 34.6 (#195)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| GPQA Diamond | 75.8% | 60.8% |
| MMLU-Pro | 79.5% | 74% |
| GPQA (HELM) | 68.4% | 59.4% |
| LMArena Expert | 1356 | 1258 |
| Confabulations | 15.7% | — |
| Vectara Hallucination Rate | 14.2% | — |

## Multilingual

- gpt-oss-120b: 48.0 (#147)
- gpt-oss-20b: 42.2 (#197)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| LMArena Non-English | 1351 | 1268 |
| LMArena Chinese | 1385 | 1314 |
| LMArena German | 1353 | 1255 |
| LMArena Japanese | 1331 | 1244 |
| LMArena Korean | 1282 | 1236 |
| LMArena Russian | 1343 | 1278 |
| LMArena Spanish | 1389 | 1267 |
| LMArena French | 1369 | — |

## Instruction Following

- gpt-oss-120b: 69.3 (#173)
- gpt-oss-20b: 61.8 (#240)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| IFEval | 83.6% | 73.2% |
| LMArena Instruction Following | 1318 | 1236 |

## Long Context

- gpt-oss-120b: 31.4 (#278)
- gpt-oss-20b: 37.9 (#209)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| LMArena Longer Query | 1319 | 1250 |
| Fiction.LiveBench | 44.4% | — |

## Writing & Preference

- gpt-oss-120b: 46.5 (#217)
- gpt-oss-20b: 35.5 (#265)

| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| LMArena Text | 1365 | 1287 |
| LMArena Creative Writing | 1275 | 1201 |
| EQ-Bench Creative Writing | 961 | 666 |
| WildBench | 84.5% | 73.7% |
| LMArena Multi-Turn | 1340 | 1268 |
| Short-Story Creative Writing | 77.1% | — |

## FAQ

### Is gpt-oss-120b better than gpt-oss-20b?

gpt-oss-120b is the stronger model overall, scoring 36.3 to 32.5 on the Noometry Index. gpt-oss-20b costs 2.0× less per token, which makes it the better buy when gpt-oss-120b's lead doesn't matter for your workload.

### Which is cheaper, gpt-oss-120b 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-oss-120b lists at $0.037 and $0.17.

### Is gpt-oss-120b or gpt-oss-20b better for coding?

gpt-oss-20b scores higher on coding benchmarks: 37.6 versus 33.5 in the Noometry coding category.

### Which has the bigger context window?

Both accept 131K tokens.

### How many benchmarks do gpt-oss-120b and gpt-oss-20b share?

34 benchmarks have published results for both models. gpt-oss-120b has 48 scored results on Noometry and gpt-oss-20b has 34.