2 days ago
7
OpenAI has enabled image interpretation and tool use in latest models. (Credit: Markus Mainka/Shutterstock)
OpenAI has released two new AI models, o3 and o4-mini, as part of its o-series architecture. The launch expands the reasoning capabilities of ChatGPT across a range of user tiers and developer access points. The models are designed to perform complex tasks involving text, code, data, and image interpretation by autonomously selecting and applying integrated tools.
ChatGPT Plus, Pro, and Team users can now access o3, o4-mini, and o4-mini-high, replacing earlier models including o1, o3-mini, and o3-mini-high. Enterprise and Education customers are expected to gain access within a week. Free-tier users can access o4-mini via the ‘Think’ option in the composer. Rate limits across all plans remain unchanged, said OpenAI.
The o3 model introduces advanced reasoning capabilities, designed to handle extended analytical tasks in domains such as software development, scientific analysis, mathematics, and visual interpretation. OpenAI reported that o3 achieved higher performance on benchmarks including Codeforces, SWE-bench, and MMMU, and reduced major task errors by 20% compared to o1.
The o4-mini model is structured for lower-latency and high-throughput use, achieving top performance on benchmarks like AIME 2024 and 2025. It supports high-volume interactions and maintains efficiency across reasoning tasks in coding, mathematics, and data interpretation, claimed the company.
Both models are accessible via the Chat Completions API and Responses API. OpenAI noted that some developers may require organisation-level verification. The Responses API supports reasoning token preservation and will soon include built-in tools for web search, file analysis, and code execution.
How OpenAI o3 and o4-mini support multimodal reasoning
The models are trained to determine when and how to use tools within a conversation. This enables multi-step problem solving within a single response cycle. In one example, OpenAI noted a query on summer energy usage in California could involve retrieving utility data, generating a Python-based forecast, visualising the results, and explaining relevant factors. The models can adapt based on the data retrieved during the process.
Visual input interpretation is also supported. The models can analyse low-quality or rotated images such as whiteboard photos or textbook diagrams, and manipulate them in real time, including rotating, zooming, or transforming visuals as part of the reasoning workflow. This capability supports multimodal problem solving, integrating both text and imagery into a single process.
OpenAI confirmed it plans to launch o3-pro with full tool access in the coming weeks. Pro users will continue to have access to o1-pro until then. The company added that future models will combine the reasoning capabilities of the o-series with the conversational design of the GPT-series to support integrated tool use in natural language settings.
The development of o3 used reinforcement learning with scaled compute resources. OpenAI reported that the model’s performance improved as both training and inference-time reasoning were increased, consistent with previous findings from GPT-series development.
Separately, Bloomberg News reported that OpenAI is in discussions to acquire Windsurf, a company focused on AI-assisted software development tools, for approximately $3bn. The deal, if finalised, would represent OpenAI’s largest acquisition to date. The report noted that the terms have not been finalised and discussions remain ongoing.
