In a significant advancement for artificial intelligence capabilities, OpenAI announced the release of its latest models, o3 and o4-mini, on April 16, 2025. These groundbreaking models represent a substantial leap forward in AI reasoning abilities, particularly in their capacity to process and integrate visual information alongside text, known as multimodal reasoning. This development marks another milestone in OpenAI’s mission to create increasingly capable and versatile AI systems.
Understanding o3 and o4-mini
OpenAI has described o3 as its most advanced model yet, specifically optimized for mathematics, coding, scientific reasoning, and image comprehension. The larger model represents the pinnacle of OpenAI’s reasoning capabilities, designed to tackle complex problems requiring multi-step thinking processes.
mini is positioned as “the most efficient model in the series, combining low latency with high-quality output, full tools support, and multimodal inputs.” This smaller alternative offers impressive performance while operating faster and at lower cost, making advanced AI capabilities more accessible for a wider range of applications.
Both models build upon the foundation established by the earlier o1 series, but with significant enhancements in their underlying architecture and capabilities. These improvements enable them to engage in more sophisticated reasoning processes that better mimic human problem-solving approaches.
Key Enhancements and Features
Multimodal Reasoning
The most revolutionary aspect of these new models is their ability to “think with images.” Unlike previous models that could merely caption or describe visual inputs, o3 and o4-mini can actively incorporate visual information into their reasoning process.
“They don’t just see an image — they think with it,” OpenAI said in a statement. “This unlocks a new class of problem-solving that blends visual and textual reasoning.”
This capability allows users to share whiteboard photos, sketches, diagrams, and other visual content that the AI can then analyze, interpret, and integrate into its problem-solving process. The models can even manipulate these images—zooming, rotating, or focusing on specific elements—as part of their analysis.
Tool Integration
“For the first time, our reasoning models can independently use all ChatGPT tools — web browsing, Python, image understanding, and image generation,” OpenAI wrote. “This helps them solve complex, multi-step problems more effectively and take real steps toward acting independently.”
This seamless integration of tools enables the models to access external information, execute code, analyze visual data, and generate images all within a single session, greatly expanding their problem-solving capabilities. For developers, this means being able to build more sophisticated AI applications that can perform multiple functions without requiring constant human guidance.
Improved Accuracy
The new models demonstrate substantial improvements in reasoning accuracy across various domains. On the GPQA Diamond benchmark, which features PhD-level science questions, o3 achieved an impressive 87.7%, significantly outperforming the average human expert score of 70%.
In coding tasks, o3 scored 71.7 on the SWE-Bench Verified benchmark, outperforming its predecessor o1 by 22.8 points. It also achieved an Elo rating of 2727 on Codeforces, a competitive programming platform, placing it among the world’s top programmers.
For mathematical reasoning, o3 scored an exceptional 96.7% on the 2024 American Invitational Mathematics Exam (AIME), showcasing advanced mathematical problem-solving abilities that far exceed previous AI models.
Simulated Reasoning
Both models employ what OpenAI calls a “private chain of thought,” allowing them to work through complex problems step by step before providing a final answer. This process more closely resembles human reasoning, where intermediate steps are considered and evaluated before reaching a conclusion.
The models can employ this capability for tasks ranging from solving mathematical problems to debugging code or analyzing complex visual diagrams. Users can even adjust the level of reasoning effort—low, medium, or high—depending on the complexity of the task at hand.
Performance Benchmarks
The performance improvements of o3 and o4-mini over previous models are substantial across all major benchmarks:
- Scientific Reasoning: On the GPQA Diamond benchmark, which features PhD-level science questions, o3 achieved 87.7%, demonstrating advanced scientific reasoning capabilities.
- Mathematical Problem-Solving: In the American Invitational Mathematics Examination (AIME), o4-mini scored 99.5% when equipped with a Python interpreter, while o3 achieved 96.7%.
- Software Engineering: On the SWE-bench Verified benchmark, o3 scored 71.7, outperforming its predecessor by 22.8 points.
- Competitive Programming: On Codeforces, a platform for algorithmic competitions, o3 achieved an Elo rating of 2727, placing it among the top programmers globally.
These benchmarks highlight not just incremental improvements but transformative advancements in AI capabilities, particularly in areas requiring complex reasoning.
Implications for Businesses and Developers
The release of o3 and o4-mini opens new possibilities for businesses and developers looking to incorporate advanced AI capabilities into their products and services:
Enhanced Problem-Solving
Organizations can leverage these models to tackle complex challenges that previously required significant human expertise. From analyzing complex data visualizations to debugging sophisticated software systems, these models offer a level of reasoning that approaches or even exceeds human capabilities in certain domains.
Multimodal Applications
The ability to process and reason with visual information enables entirely new categories of applications. Engineering firms can analyze technical diagrams, healthcare organizations can interpret medical imaging alongside textual data, and educational platforms can provide more comprehensive explanations that incorporate visual elements.
Autonomous Systems
The models’ ability to use tools independently, including web browsing, code execution, and image analysis, lays the groundwork for more autonomous AI systems. As OpenAI’s Greg Brockman noted during the announcement, “There are some models that feel like a qualitative step into the future. GPT-4 was one of those. Today is also going to be one of those days.”
Development Workflows
For developers specifically, OpenAI has introduced Codex CLI, a lightweight coding agent that runs directly in a terminal. This tool allows developers to leverage the models’ reasoning capabilities for coding tasks, with support for screenshots and sketches.
Access and Availability
Both o3 and o4-mini became available to ChatGPT Plus, Pro, and Team customers starting April 16, 2025. API access is also being rolled out to developers, enabling integration into third-party applications and services.
For those using Microsoft’s Azure platform, the models are available through Azure OpenAI Service in Azure AI Foundry and GitHub, offering enterprise-grade security and compliance features.
As with previous models, access comes with usage limits based on subscription tier. ChatGPT Plus, Team, and Enterprise accounts have specific message allocations, while API users on paid tiers have access based on their service level.
Conclusion
The introduction of o3 and o4-mini represents a significant milestone in AI development, particularly in the realm of reasoning capabilities. By combining advanced multimodal understanding with improved accuracy and tool integration, these models offer unprecedented capabilities for solving complex problems across numerous domains.
As businesses and developers begin to explore the potential applications of these new models, we can expect to see innovative solutions emerging in fields ranging from software development and scientific research to education and creative industries. The ability to “think with images” opens new frontiers for human-AI collaboration, potentially transforming how we approach complex problems that involve both visual and textual elements.
While the long-term impact of these advancements remains to be seen, it’s clear that OpenAI’s o3 and o4-mini models mark another significant step toward more capable, versatile, and reasoning-oriented artificial intelligence systems.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
