0

o3

Closed

OpenAI's first true "reasoning at scale" model, announced Dec 2024 and publicly released April 2025, which crossed human-expert ceiling on GPQA.

Open
Publisher
OpenAI
Family
O Series
Params
Undisclosed (closed)
Context
200K
Input
Output
Providers
openai
License
Proprietary
Released
Apr 2025
Canonical
openai.com/index/introducing-o3-and-o4-mini

Cite

Notes

Only stored in your browser.

Attribution

Attribution policy →
Context
200K
Input
$2
/1M tok
Output
$8
/1M tok
Blended
$3.50
#369/ 414/1M · 3:1
Speed
137
#70/ 414tok/s
Cache read
$0.50
0.25x input

Price & speed via Artificial Analysis · ranked across tracked models · caching via OpenRouter

Reported on 27 evals across 11 domains - best 99.2% on MATH-500 · #1 of 178

Reported eval scores

27
MATH-50099.2
AIME 2024: Problems from the American Invitational Mathematics Examination96.7
FrontierMath18.7
Arena-Hard88.8
MMLU-Pro85.3
Humanity's Last Exam (HLE)20
SWE-bench Multimodal36
IFBench71.4
Terminal-Bench (Hard)37.1
GPQA Diamond87.7
SciCode41
LiveCodeBench80.8
Aider Polyglot Benchmark81.3
SWE-bench Verified58.4
SWE-bench71.7
AIME 2025: Problems from the American Invitational Mathematics Examination88.3
AIME202491.6
AIME202588.9
TaxEval v274.6
MortgageTax65.7
CorpFin v259.7
τ²-bench (Tau²-bench)80.7
BrowseComp0
OSWorld-Verified23
MedCode47.3
MedScribe76.7
SAGE41.8