OpenAI to Retire o3 and GPT-4.5 Models

OpenAI o3 Model:

• Architecture: Reflective Generative Pre-trained Transformer (GPT) model, successor to OpenAI o1.
• Reasoning Mechanism: Designed to devote additional deliberation time for step-by-step logical reasoning, utilizing a 'private chain of thought' and trained with large-scale reinforcement learning on chains of thought.
• Capabilities: Excels in complex tasks like coding, mathematics, science, and visual perception.
• Features: Offers adjustable reasoning effort levels (low, medium, high) to balance speed and thoroughness. Integrates full tool capabilities including web browsing, Python code execution, file analysis, image generation, canvas, automations, file search, and memory.
• Variants: Includes o3-mini (cost-efficient, specialized for technical domains requiring precision and speed) and o3-pro (most capable, designed for deep, logical thinking).

GPT-4.5 Model:

• Codenamed: "Orion".
• Architecture: Large language model built upon a modified transformer structure.
• Training Methodology: Primarily trained using unsupervised learning to enhance pattern recognition, draw connections, and generate creative insights without explicit reasoning. This was combined with supervised fine-tuning and reinforcement learning from human feedback.
• Computational Resources: Trained on Microsoft Azure AI supercomputers.
• Key Architectural Innovations: Dynamic attention routing (conditional computation pathways), hierarchical token processing (64k context windows with tiered attention mechanisms), and multi-modal fusion layers for integrating text, image, and code representations.
• Training Data: Utilized a pre-training corpus of 13.8 trillion tokens from diversified sources.
• Capabilities: Known for broader knowledge, deeper world understanding, reduced hallucinations, and improved reliability. Supports vision capabilities through image inputs.
• Cost: Described by Sam Altman as a "giant, expensive model."