On Wednesday, OpenAI announced two upgraded AI models — OpenAI o3 and OpenAI o4-mini — that can now understand and work with images and text together. That means they can analyze sketches, posters, diagrams, and graphs, not just written content.
What Can These New OpenAI Models Do?
According to OpenAI's Head of Research, Mark Chen, these systems can manipulate, crop, and transform images as part of solving a task. They’re not just passively analyzing content — they’re actively working with it.
These new AI models also come with powerful capabilities:
-
Generate images
-
Search the web
-
Use digital tools to assist with more complex tasks
Unlike earlier versions of ChatGPT, which gave quick answers, these new models “think” through tasks, solving them step-by-step — more like how humans reason.
Why This Matters for Developers and Tech Users
OpenAI’s reasoning systems are based on large language models (LLMs). These are trained using a process called reinforcement learning, where the AI learns by trial and error, figuring out the best strategies over time.
This kind of AI is especially useful for developers and programmers. With reasoning ability, the AI can:
-
Help debug code
-
Suggest better coding solutions
-
Understand complex logic
-
Solve math and programming problems visually
Introducing Codex CLI: A New AI Tool for Programmers
Alongside o3 and o4-mini, OpenAI also released a new tool called Codex CLI, an AI agent designed to help programmers work with code stored on their computers.
Codex CLI can use these new AI models to understand, edit, and improve code. OpenAI has open-sourced this tool, so developers can customize and build their own solutions using it.
Who Can Use These New Tools?
These new AI models — o3 and o4-mini — are now available to subscribers of:
-
ChatGPT Plus ($20/month)
-
ChatGPT Pro ($200/month), which gives full access to all the latest tools from OpenAI
Final Thoughts: A New Era of Visual and Text-Based AI
OpenAI’s latest update is part of a larger race to develop AI systems that reason like humans. Tech giants like Google and Meta, along with startups like DeepSeek, are working on similar models.
While these tools are powerful, experts caution that they’re not perfect — they can still make mistakes, or even “hallucinate” answers (produce incorrect or made-up information).