gpt-oss-20b

Across the task benchmarks we track, gpt-oss-20b’s strongest relative results are in knowledge work, terminal use, and reasoning. On GDPval-AA v2.1, it scores 347 Elo, ranking #37 among 38 models with reported results in our dataset.

Updated: Nov 5, 2026Data as of: Nov 2, 2026

Compare this model

Strongest results

Strongest results for gpt-oss-20b in tracked evaluation cohorts
Benchmark	Measures	Score
Terminal-Bench ↗4.0 / mini-swe-agent	Working in a terminal	0%Joint #49 of 56
Humanity's Last Exam ↗Stable HLE — text-only subset	Expert reasoning	10.98%#63 of 68
IFBench ↗58 constraint types	Precise instruction following	65.1%#16 of 17
HLE-Diamond ↗1,000 questions / no tools	Advanced knowledge and reasoning	4.8%Joint #33 of 35

Model details

Open source / open weights11 benchmarks

Specifications for gpt-oss-20b
Model	gpt-oss-20b
Agent Index v1.2	Insufficient coverage (4/7) · Preview
Creator	OpenAI (opens in a new tab)
Availability	Open source / open weights
Context size	131,072 tokens (opens in a new tab)
Max output	Not reported
Knowledge cutoff	Not reported
License	Apache 2.0 · Model weights ↗ (opens in a new tab)

Pricing and performance

Per provider and reported configuration. Prices are USD per 1M tokens.

No provider pricing or performance reports have been added for this model yet.

Benchmark performance

gpt-oss-20b in color, alongside leading models. Latest stored reports, best effort per report date.

Reasoning

198-question Diamond subset; independent AA evaluation

GPQA Diamond ↗

01
GPT-6 Astra96.3%
02
Gemini 3.8 Flash95.3%
03
Gemini 3.1 Pro94.1%
03
GPT-5.6 Sol94.1%
03
Muse Spark 1.394.1%
30
gpt-oss-20bThis model68.8%

Terminal Use

89 tasks; Terminus 2; e2b; three repeats

Terminal-Bench 2.1 ↗

01
Kimi K385%
02
GLM-5.3-Flash84.3%
03
GLM-5.383.9%
04
DeepSeek V4 Pro 081378.7%
05
GLM-5.277.9%
13
gpt-oss-20bThis model13.9%

Work Automations

Private 657 tasks; 50 turns; objective completion

AutomationBench-AA ↗

01
Gemini 4 Carbon77.8%
02
Gemini 4 Argon77.5%
03
Kimi K3.1 Pro73.3%
04
MiniMax M3.1 Flash72.52%
05
Claude Sonnet 5.571.8%
55
gpt-oss-20bThis model0.19%

Knowledge Work

220 tasks; Stirrup agent; pairwise Elo

GDPval-AA v2.1 ↗

01
Claude Opus 5.51,866 Elo
02
Kimi K3.1 Pro1,858 Elo
03
Claude Sonnet 5.51,838 Elo
04
Claude Fable 5.11,763 Elo
05
DeepSeek V4.1 Flash 10171,746 Elo
37
gpt-oss-20bThis model347 Elo

Reported benchmarks

All published results in our dataset, with versions, evaluation settings, and sources.

All published benchmark reports for gpt-oss-20b
Benchmark	Score	Evaluation setup	Source
AA-Briefcase1.1 / combined EloCombined Elo · Higher is better	0 Elo	selection: best reported configuration per release · gpt-oss-20b (High) · tools: benchmark-defined	Artificial Analysis (opens in a new tab)independent
AppWorldTest-ChallengeTask Goal Completion · Higher is better	15.3%	agent: RooAGI_agentbuilder_memtune_rl · metric: TGC · selection: All 23 leaderboard submissions; latest report date and best score at that date shown per model. Normal-split and SGC values retained in the source data/settings, not mixed into the ranking. · source model label: gptoss-20b · split: Test-Challenge · test challenge sgc: 7.9 · test normal sgc: 58.9 · test normal tgc: 76.2	AppWorld research team ↗ (opens in a new tab)benchmark publisher · 2026-02-22
AutomationBench-AA1.0.6 / AA private 657Objectives completed without guardrail violations · Higher is better	0.191%	gpt-oss-20b (High); Private 657 tasks; 50 turns; objective completion	Artificial Analysis (opens in a new tab)independent
GDP.pdf100 tasks / All-pass / five attemptsAccuracy · Higher is better	2%	selection: best reported configuration per release · gpt-oss-20b (High) · tools: benchmark-defined	Artificial Analysis (opens in a new tab)independent
GDPval-AA2.1Elo · Higher is better	347.19 Elo	gpt-oss-20b (High); 220 tasks; Stirrup agent; pairwise Elo	Artificial Analysis (opens in a new tab)independent
GPQADiamond / Artificial AnalysisAccuracy · Higher is better	68.788%	gpt-oss-20b (High); 198-question Diamond subset; independent AA evaluation	Artificial Analysis (opens in a new tab)independent
GPQADiamond / Artificial AnalysisAccuracy · Higher is better	61.111%	gpt-oss-20b (Low); 198-question Diamond subset; independent AA evaluation	Artificial Analysis (opens in a new tab)independent
HLE-Diamond1,000 questions / no toolsAccuracy · Higher is better	4.8%	selection: best reported configuration per release · gpt-oss-20b (High) · No tools	Artificial Analysis (opens in a new tab)independent
Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better	10.982%	gpt-oss-20b (High); original HLE text-only 2,158 questions; no tools; not HLE Diamond · No tools	Artificial Analysis (opens in a new tab)independent
Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better	5.283%	gpt-oss-20b (Low); original HLE text-only 2,158 questions; no tools; not HLE Diamond · No tools	Artificial Analysis (opens in a new tab)independent
IFBench58 constraint typesAccuracy · Higher is better	65.102%	selection: best reported configuration per release · gpt-oss-20b (High) · tools: benchmark-defined	Artificial Analysis (opens in a new tab)independent
Terminal-Bench2.1 / Artificial AnalysisPass@1 · Higher is better	13.858%	gpt-oss-20b (High); 89 tasks; Terminus 2; e2b; three repeats	Artificial Analysis (opens in a new tab)independent
Terminal-Bench4.0 / mini-swe-agentTask success · Higher is better	0%	harness: mini-swe-agent · repeats: 3 · selection: best reported configuration per release · gpt-oss-20b (High)	Artificial Analysis (opens in a new tab)independent