OpenAI Releases New AI Models: o3 and o4 Mini
OpenAI continues to innovate rapidly in the field of artificial intelligence. Following the recent unveiling of GPT-4.1, the company has officially launched two new AI models: o3 and o4 mini. These models promise enhanced capabilities and accessibility for users.
Key Features of o3 and o4 Mini
These new models build upon the foundation of GPT-4o, incorporating its multimodal capabilities. Importantly, they offer full tool access, including:
-
Image generation
-
Web browsing
-
Code execution
A groundbreaking feature is the introduction of AI with the ability to "think" about images, further blurring the lines between human and artificial intelligence.
Model Options and Performance
OpenAI users will now find three models available upon logging in:
-
o3: Designed for advanced reasoning.
-
o4 mini: Suitable for quick and advanced reasoning.
-
o4 mini High: Excels in coding and visual reasoning.
Benchmark results demonstrate that the latest versions of o3 and o4 mini surpass the older o1 model in various areas, including coding, mathematical problem-solving, and human preference evaluations. Both the regular and "High" versions of o3 and o4 mini outperform their predecessor.
Testing Advanced Reasoning with the Einstein Riddle
To test the advanced reasoning capabilities of these models, the Einstein Riddle was used. The o3 model was able to correctly identify the person who owns the fish based on the provided clues, demonstrating its effective use of the process of elimination.
Evaluating Reasoning Speed with a Time-Based Question
The models were also tested with a time-based question to evaluate their reasoning speed. The o4 mini model was able to correctly answer "Friday".
Solving a Mathematical Puzzle with o4 Mini
To test the model's mathematical prowess, the o4 mini was given a puzzle to identify a fake coin. The correct answer is three weighings. The o4 mini was able to get the correct answer and explain the logic with three possible outcomes.
Visual Reasoning Capabilities with o4 Mini High
The o4 mini High model, specializing in coding and visual reasoning, was tested by providing it with a game screenshot.
-
The model was instructed to create a similar air combat game using p5js.
-
The model successfully generated code that created a working game resembling the screenshot.
-
The initial code was then refined based on further instructions, improving the game's visuals.
The speed at which the o4 mini High model generates code is particularly noteworthy, easily handling requests for complex games like a Super Mario Bros. clone. It even added a coin!
Testing Image Interpretation
-
The system correctly answered a puzzle by analyzing the L-shaped image to select the answer with three.
-
The model was asked what a five-year-old's drawing was trying to portray. The model was able to state it was the father and family and correctly guessed the person in the image was the father.
Image Generation
The image generation ability of o4 mini high was tested by asking it to create a unicorn anime girl. The mode was able to generate the image quickly.
Enhanced Web Search and Data Visualization
The models also exhibit improved web search functionality. For example:
-
The model was tasked with analyzing the market share of Windows 11, Windows 10, macOS, and Linux operating systems.
-
It successfully gathered data from multiple websites and presented the information in an interactive chart.
-
The user can customize the chart's colors and download it.
This feature is particularly useful for creating data visualizations without needing specialized software or personnel.
Accessing o3 and o4 Mini for Free
For users without a paid OpenAI membership, alternative methods exist to access the o3 and o4 mini models:
- Download the Cursor client.
- Download Windsurf.
These clients offer free usage allowances, allowing users to explore the capabilities of the new models.
Using the Cursor Client
- Download and install the Cursor client, choosing the appropriate version for your operating system (Windows, macOS, or Linux).
- Create a free account or log in if you already have one.
- Navigate to the settings and enable the o3 and o4 mini models.
- Select the desired model and begin using it.