OpenAI logo

o4-mini

Multimodal
OpenAI

o4-mini is OpenAI's latest small model in the o-series, optimized for fast and efficient reasoning with exceptionally high performance on coding and visual processing tasks. It operates faster and costs less than o3.

Key Specifications

Parameters
-
Context
200.0K
Release Date
April 16, 2025
Average Score
66.5%

Timeline

Key dates in the model's history
Announcement
April 16, 2025
Last Update
July 19, 2025
Today
March 25, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
May 31, 2024
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$1.10
Output (per 1M tokens)
$4.40
Max Input Tokens
200.0K
Max Output Tokens
100.0K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Programming

Programming skills tests
SWE-Bench Verified
accuracySelf-reported
68.1%

Reasoning

Logical reasoning and analysis
GPQA
accuracy (without tools)Self-reported
81.4%

Multimodal

Working with images and visual data
MathVista
AccuracySelf-reported
84.3%
MMMU
accuracySelf-reported
81.6%

Other Tests

Specialized benchmarks
Aider-Polyglot
accuracy (all sample, o4-mini-high)Self-reported
68.9%
Aider-Polyglot Edit
Accuracy (diff, o4-mini-high)Self-reported
58.2%
AIME 2024
Accuracy (without tools)Self-reported
93.4%
AIME 2025
accuracy (without tools)Self-reported
92.7%
BrowseComp
Accuracy (with Python + in ) AI: I I will how can answer on questions at testing, using Python (if necessary) and search in (if necessary) for improvement accuracy. At is access to: - Python for computations, analysis data and mathematical tasks - in for obtaining information I I will: 1. Python for analysis and solutions tasks, requiring programming 2. search in for search facts, and information 3. when I tools 4. exact, answers with 5. code, output and in answers I not I will: 1. if not 2. information 3. "accuracy" (that I when this not so) 4. search or code, when I I can answer without them goal — accuracy at each answerSelf-reported
51.5%
CharXiv-R
accuracySelf-reported
72.0%
Humanity's Last Exam
accuracy (without tools)Self-reported
14.7%
Scale MultiChallenge
accuracySelf-reported
43.0%
TAU-bench Airline
accuracy (o4-mini-high)Self-reported
49.2%
TAU-bench Retail
Accuracy (o4-mini-high)Self-reported
71.8%

License & Metadata

License
proprietary
Announcement Date
April 16, 2025
Last Updated
July 19, 2025

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.