OpenAI o3 and o4-mini release notes / what's new

OpenAI has released o3 and o4-mini, the latest models in the o-series designed for extended reasoning. These models represent a significant advancement in intelligence and utility by integrating state-of-the-art reasoning with the ability to agentically use and combine tools—including web search, Python data analysis, and image generation—to solve multi-faceted problems typically in under a minute.

Frontier Reasoning with OpenAI o3

OpenAI o3 is the most powerful reasoning model released to date, establishing new state-of-the-art (SOTA) performance on benchmarks including Codeforces, MMMU, and SWE-bench (evaluated on a fixed subset of 477 verified tasks without a custom model-specific scaffold).

Key performance improvements include:

  • Error Reduction: In external expert evaluations, o3 makes 20% fewer major errors than OpenAI o1 on difficult real-world tasks, particularly in creative ideation, business/consulting, and programming.
  • Analytical Rigor: Early testers noted o3's ability to generate and critically evaluate novel hypotheses in engineering, math, and biology contexts.
  • Visual Perception: o3 performs strongly in analyzing graphics, charts, and images.

Cost-Efficient Reasoning with OpenAI o4-mini

OpenAI o4-mini is a smaller, optimized model designed for high-volume, high-throughput reasoning tasks. It outperforms its predecessor, o3-mini, in data science and non-STEM domains.

Notable benchmarks include:

  • AIME Performance: o4-mini is the best-performing benchmarked model on AIME 2024 and 2025. When provided with a Python interpreter, it achieves a 99.5% pass@1 (100% consensus@8) on AIME 2025.
  • Efficiency: Due to its size, o4-mini supports significantly higher usage limits than o3 while maintaining strong performance in coding, math, and visual tasks.

Multimodal Reasoning and Visual Integration

For the first time, o-series models integrate images directly into their chain of thought, allowing them to "think with images" rather than simply perceiving them. This enables the models to solve complex problems by blending visual and textual reasoning.

Capabilities include:

  • Robust Interpretation: The models can interpret blurry, low-quality, or reversed images, such as hand-drawn sketches, textbook diagrams, or whiteboard photos.
  • Active Manipulation: Through tool use, the models can zoom, rotate, or transform images as part of the reasoning process to better analyze the input.

Agentic Tool Use and Workflow Automation

OpenAI o3 and o4-mini are trained via reinforcement learning to reason about when and how to use tools to achieve a desired outcome. They have full access to ChatGPT tools and custom tools via API function calling.

This strategic approach allows for multi-step workflows, such as:

  • Iterative Research: Searching the web multiple times, analyzing results, and pivoting search queries based on discovered information.
  • Complex Synthesis: Chaining web search for public data, Python for forecasting, and image generation for visualization to answer a single complex query.

Scaling Reinforcement Learning (RL)

OpenAI observed that large-scale RL follows a "more compute = better performance" trend similar to GPT-series pretraining. By increasing training compute and inference-time reasoning by an order of magnitude, OpenAI has validated that performance continues to climb as models are allowed to "think" longer.

At equal latency and cost to OpenAI o1, o3 delivers higher performance, with further gains possible by increasing the reasoning time.

Safety and Risk Mitigation

OpenAI rebuilt the safety training data for these models, introducing new refusal prompts for jailbreaks, malware generation, and biological threats (biorisk).

Safety measures include:

  • Reasoning LLM Monitor: A monitor trained on human-written safety specifications that flagged approximately 99% of conversations in biorisk human red-teaming campaigns.
  • Preparedness Framework: Both models were evaluated across cybersecurity, AI self-improvement, and biological/chemical risks. Both o3 and o4-mini remain below the "High" threshold in all three categories.

Codex CLI and Developer Access

OpenAI introduced Codex CLI, an open-source lightweight coding agent that runs in the terminal. It allows users to pass screenshots or sketches to o3 and o4-mini while providing the model access to local code.

Access Details:

  • ChatGPT: Plus, Pro, and Team users have access to o3, o4-mini, and o4-mini-high. Enterprise and Edu users gain access one week later. Free users can access o4-mini via the 'Think' option.
  • API: Both models are available via Chat Completions and Responses APIs. The Responses API supports reasoning summaries and the preservation of reasoning tokens around function calls.
  • o3-pro: A version of o3 designed for longer thinking and higher reliability is available to Pro users in ChatGPT and the API (as of June 10, 2025).

Sources