Skip to content

OpenAI

o3

A preview set a new high score on ARC-AGI-1 in December 2024. ARC Prize says the o3 released in April 2025 is not the version it tested.

Announced
Released
Access
Public

Benchmark scores

TestScoreBest on the trackerHow it was measured
Epoch Capabilities Indexmany tests combined into one score147 pointsGPT-6.2 Astra 174 pointsIndependentlisted as "o3", CI 145.54 to 148.54Run 9 Nov 2026
Humanity's Last Examexpert questions, no tools20%GPT-6 Astra 55%Independentleaderboard entry "o3 (high) (April 2025)" (o3-2025-04-16_high)
FrontierMath T1–3hard maths problems33%Claude Fable 5.2 95%Independenthigh reasoning effort (o3-2025-04-16_high)Run 27 Aug 2026
OSWorld-Verifiedusing a computerPeople (original task set): 72%23%Claude Fable 5 86%Independentplain model, max 100 steps ("o3", 83.03/361 tasks)Run 28 Jul 2025
GPQA DiamondPhD-level science questionsPhD experts: 70%Saturated82%Gemini 4 Argon 96%Independenthigh reasoning effort (o3-2025-04-16_high)Run 16 Apr 2025
ARC-AGI-2visual puzzles, model aloneAverage person: 60%Saturated6.5%GPT-6 Astra 95%IndependentARC Prize config "o3 (High)" (o3-2025-04-16-high)
OTIS Mock AIMEcompetition mathsSaturated84%GPT-5.5 100%Independentmedium reasoning effort (o3-2025-04-16_medium)Run 7 Aug 2026
SWE-bench Verifiedfixing real software bugsKnown flaws62%Claude Opus 4.7 83%Independentmedium reasoning effort (o3-2025-04-16_medium)Run 12 Feb 2026
METR time horizontask length an AI agent finishes half the timeNo longer updated2 hClaude Mythos Preview 17 hIndependentMETR Horizon v1.1 run 'o3_inspect'; 95% CI 74.6–190.9 min

Compare every model on the same tests →

Also on these charts

Sources