gpt-oss-120b vs gpt-oss-20b
Wins 6 of 6 areas
Coding · Agents · Reasoning · Facts · Long documents · Following instructions
Wins 0 of 6 areas
—
gpt-oss-120b is the stronger all-rounder.gpt-oss-20b is cheaper.
Scores updated · 27 tests both models report · How we compare
Where each one wins
Tests won in each of the six areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- AgentsCarrying out multi-step tasks on its own20gpt-oss-120b2 of 3 tests · 1 tie
- ReasoningHard problems that need careful thinking20gpt-oss-120b2 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up20gpt-oss-120b2 of 2 tests
- CodingWriting and fixing software21gpt-oss-120b2 of 3 tests
- Long documentsFinding answers in very long texts10gpt-oss-120b1 of 1 test
- Following instructionsDoing exactly what it is asked10gpt-oss-120b1 of 1 test
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
gpt-oss-20b costs 66% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where gpt-oss-120b pulls ahead
- Reasons across sets of long documentsAA-LCR+17.3points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+12.9points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+12.3points ahead
Where gpt-oss-20b pulls ahead
- Code for real scientific research problemsSciCode+4.9points ahead
Every test, side by side
All 27 tests both models report. The winning score is in its model's colour; marks a score checked independently.
Codinggpt-oss-120b
- Terminal-Bench Hardgpt-oss-120b by 12.923.510.6+12.9
- Terminal-Bench 2.1gpt-oss-120b by 12.326.213.9+12.3
- SciCodegpt-oss-20b by 4.93438.9+4.9
Agentsgpt-oss-120b
- τ-Bench V3 · Bankinggpt-oss-120b by 5.812.87+5.8
- GDPValgpt-oss-120b by 5.65.60+5.6
- Terminal-Bench 4.0tie00tie
Reasoninggpt-oss-120b
- GPQA Diamondgpt-oss-120b by 9.478.268.8+9.4
- Humanity's Last Examgpt-oss-120b by 8.619.611+8.6
- CritPttie1.11.4tie
Factsgpt-oss-120b
- AA-Omniscience · Accuracygpt-oss-120b by 5.821.816+5.8
- AA-Omniscience · Non-hallucinationgpt-oss-120b by 3.39.25.9+3.3
Following instructionsgpt-oss-120b
- IFBenchgpt-oss-120b by 3.96965.1+3.9
Other results14 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- HMMT Feb. 2025gpt-oss-120b by 2096.776.7+20
- HealthBenchgpt-oss-120b by 15.157.642.5+15.1
- AA-Omnisciencegpt-oss-120b by 13.8-49.3-63+13.8
- TAU-bench (retail)gpt-oss-120b by 1367.854.8+13
- Artificial Analysis Coding Indexgpt-oss-120b by 9.730.420.7+9.7
- τ²-Bench Telecom (AA run)gpt-oss-120b by 5.665.860.2+5.6
- XSTestgpt-oss-120b by 5.592.787.2+5.5
- AA Intelligencegpt-oss-120b by 2.611.69+2.6
- AIR-Bench 2024gpt-oss-120b by 28886+2
- bbqgpt-oss-120b by 1.898.596.7+1.8
- HarmBenchgpt-oss-120b by 1.310098.7+1.3
- AIME 2025tie9089.2tie
- anthropic_red_teamtie99.599.7tie
- simple_safety_teststie100100tie
Questions people ask
Which is better, gpt-oss-120b or gpt-oss-20b?
gpt-oss-120b wins all six areas where both have results: coding, agents, reasoning, facts, long documents and following instructions. gpt-oss-20b wins none, but costs 66% less.
Which is better for coding?
gpt-oss-120b. It wins 2 of the 3 coding tests both models report; gpt-oss-20b wins 1.
Which is cheaper?
gpt-oss-120b costs $0.15 per million input tokens and $0.59 per million output tokens; gpt-oss-20b costs $0.07 and $0.18. That makes gpt-oss-20b about 66% cheaper for the same work.
How do you compare the two?
We use the 27 benchmark tests both models have published scores on. The verdict counts the 13 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 14 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.