Model comparison
gpt-oss-20b vs o4-mini
o4-mini is the stronger model overall, scoring 41.6 to 32.5 on the Noometry Index. gpt-oss-20b costs 53× less per token, which makes it the better buy when o4-mini's lead doesn't matter for your workload.
Last verified . 31 shared benchmarks.
Summary
- They share 31 benchmarks with published results for both. gpt-oss-20b scores higher in 0 categories and o4-mini in 9 categories; 9 gaps are clear of the uncertainty.
- The widest gap is in agentic & tool use, where o4-mini leads 32.6 to 9.3.
- The biggest single-benchmark swing is Chess Puzzles: 4% for gpt-oss-20b and 26% for o4-mini.
- gpt-oss-20b is cheaper at $0.018 / $0.09 per million input/output tokens, against $1.10 / $4.40 for o4-mini.
- o4-mini accepts more context: 200K tokens versus 131K.
- gpt-oss-20b has downloadable open weights; the other is API-only.
Side by side
| gpt-oss-20b | o4-mini | |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Noometry Index | 32.5 | 41.6 |
| Released | 2025-08-05 | 2025-04-16 |
| Weights | Open | Proprietary |
| Context window | 131K | 200K |
| Max output | 16K | 100K |
| Input $ / M tokens | $0.018 | $1.10 |
| Output $ / M tokens | $0.09 | $4.40 |
| Results tracked | 34 | 60 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding o4-mini leads
gpt-oss-20b: 37.6 (#192), o4-mini: 40.9 (#127)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| WeirdML | 40.9% | 52.6% |
| LMArena Coding | 1306 | 1368 |
| ALE-Bench | 566.05 | 826.17 |
| SWE-bench Verified (bash only) | — | 45% |
| Aider Polyglot | — | 72% |
| SciCode | 34.4% | — |
| GSO | — | 3.6% |
| CadEval | — | 62% |
| AlgoTune | — | 1.72 |
Agentic & Tool Use o4-mini leads
gpt-oss-20b: 9.3 (#154), o4-mini: 32.6 (#61)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| Terminal-Bench | 3.4% | — |
| Berkeley Function Calling Leaderboard | — | 53.2% |
| GDPval | — | 25.3% |
| METR Time Horizons | — | 63.9% |
Reasoning o4-mini leads
gpt-oss-20b: 19.3 (#261), o4-mini: 24.6 (#162)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| Kagi LLM Benchmark | 53.2% | 67.6% |
| CritPt | 1.4% | 0.6% |
| Chess Puzzles | 4% | 26% |
| LMArena Hard Prompts | 1274 | 1351 |
| DTBench | 68% | 77.6% |
| LMCA | 14.5% | 26.5% |
| Epoch Capabilities Index | 137.82 | 145.64 |
| ARC-AGI-2 | — | 6.1% |
| SimpleBench | — | 38.7% |
| ARC-AGI-1 | — | 58.7% |
| EnigmaEval | — | 9.2% |
| Mystery Game Puzzles | — | 5% |
| ForecastBench | — | 61.8 |
Math o4-mini leads
gpt-oss-20b: 39.4 (#103), o4-mini: 40.8 (#89)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 65.3% | 81.7% |
| Omni-MATH | 56.5% | 72% |
| LMArena Math | 1317 | 1389 |
| FrontierMath (Tiers 1-3) | — | 36.1% |
| FrontierMath Tier 4 | — | 4.9% |
| MATH Level 5 | — | 97.8% |
| FrontierMath (Feb 2025 set) | — | 24.8% |
| FrontierMath Tier 4 (v1) | — | 6.3% |
Knowledge o4-mini leads
gpt-oss-20b: 34.6 (#195), o4-mini: 43.6 (#91)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| GPQA Diamond | 60.8% | 79.6% |
| MMLU-Pro | 74% | 82% |
| GPQA (HELM) | 59.4% | 73.5% |
| LMArena Expert | 1258 | 1343 |
| Humanity's Last Exam | — | 18.1% |
| SimpleQA Verified | — | 19.6% |
| Confabulations | — | 15.8% |
| Vectara Hallucination Rate | — | 18.6% |
Multimodal Not comparable
gpt-oss-20b: —, o4-mini: 40.2 (#49)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| LMArena Vision | — | 1194 |
| GeoBench | — | 64% |
| VPCT | — | 57.5% |
Multilingual o4-mini leads
gpt-oss-20b: 42.2 (#197), o4-mini: 47.0 (#154)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| LMArena Non-English | 1268 | 1337 |
| LMArena Chinese | 1314 | 1354 |
| LMArena German | 1255 | 1336 |
| LMArena Japanese | 1244 | 1308 |
| LMArena Korean | 1236 | 1312 |
| LMArena Russian | 1278 | 1334 |
| LMArena Spanish | 1267 | 1347 |
| LMArena French | — | 1364 |
Instruction Following o4-mini leads
gpt-oss-20b: 61.8 (#240), o4-mini: 75.2 (#68)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| IFEval | 73.2% | 92.8% |
| LMArena Instruction Following | 1236 | 1321 |
Long Context o4-mini leads
gpt-oss-20b: 37.9 (#209), o4-mini: 45.5 (#33)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| LMArena Longer Query | 1250 | 1315 |
| Fiction.LiveBench | — | 77.8% |
Writing & Preference o4-mini leads
gpt-oss-20b: 35.5 (#265), o4-mini: 54.0 (#152)
| Benchmark | gpt-oss-20b | o4-mini |
|---|---|---|
| LMArena Text | 1287 | 1353 |
| LMArena Creative Writing | 1201 | 1294 |
| WildBench | 73.7% | 85.4% |
| LMArena Multi-Turn | 1268 | 1350 |
| Short-Story Creative Writing | — | 75% |
| EQ-Bench Creative Writing | 666 | — |
Frequently asked questions
Is gpt-oss-20b better than o4-mini?
o4-mini is the stronger model overall, scoring 41.6 to 32.5 on the Noometry Index. gpt-oss-20b costs 53× less per token, which makes it the better buy when o4-mini's lead doesn't matter for your workload.
Which is cheaper, gpt-oss-20b or o4-mini?
gpt-oss-20b is cheaper. It lists at $0.018 per million input tokens and $0.09 per million output tokens; o4-mini lists at $1.10 and $4.40.
Is gpt-oss-20b or o4-mini better for coding?
o4-mini scores higher on coding benchmarks: 40.9 versus 37.6 in the Noometry coding category.
Which has the bigger context window?
o4-mini does, with 200K tokens against 131K.
How many benchmarks do gpt-oss-20b and o4-mini share?
31 benchmarks have published results for both models. gpt-oss-20b has 34 scored results on Noometry and o4-mini has 60.