Introducing GPT-4.5 Preview

OpenAI has just announced the launch of GPT-4.5-preview, the latest and largest model available as a research preview, marking a significant advancement in artificial intelligence capabilities. This iteration enhances the understanding of user intent, enabling more natural, nuanced conversations. Improvements in creativity, empathy, and general knowledge make GPT-4.5 ideal for a variety of applications including writing, brainstorming, coaching, and planning.

Key Features and Advancements

GPT-4.5 brings forth a range of new features designed to facilitate smoother AI interactions. Among these enhancements are:

• Support for function calling: This allows the model to interface more effectively with external systems.
• Structured outputs: It offers organized response formats, making it easier for users to digest information.
• Vision support: This capability allows the model to understand and analyze visual inputs, expanding its usability.
• Streaming and system messages: Users can receive dynamic updates, improving the interactivity of sessions.
• Evals and prompt caching: These features streamline processes, particularly in frequent or repetitive interactions.

Access to the model is facilitated via an API supporting Chat Completions, Assistants, and Batch APIs. While GPT-4.5 is more computationally intensive and larger than its predecessor, GPT-4o, it is not a direct replacement but rather a complementary evolution.

In terms of cost structure, the pricing is as follows:

• $68 per 1M tokens (average query)
• $75 per 1M input tokens
• $37.5 per 1M cached input tokens
• $150 per 1M output tokens
• Batch jobs and cached input are discounted by 50%

Scaling AI: Unsupervised Learning vs. Reasoning

GPT-4.5 makes strides in AI development through two pivotal paradigms:

• Scaling reasoning: This model excels in step-by-step problem-solving across STEM and logic domains.
• Scaling unsupervised learning: This enhancement broadens the model’s general world knowledge, reducing hallucinations and improving reliability.

Key to these developments is the advanced training infrastructure, operationalized on Microsoft Azure AI supercomputers, leveraging innovation in both architecture and optimization techniques for significantly improved performance.

Performance Improvements Over Previous Models

GPT-4.5 demonstrates marked improvements in factual accuracy and a reduction in hallucinations. In SimpleQA accuracy tests, the model outperformed GPT-4o and other OpenAI models. Its enhanced emotional intelligence facilitates better human collaboration, enabling a deeper understanding of nuanced conversations.

Comparative evaluations show that GPT-4.5 earns a higher win rate in creative intelligence tasks, professional queries, and everyday conversations than its predecessor, proving its capacity for versatile engagement.

Real-World Use Cases

The use cases for GPT-4.5 are vast and promising:

• Personal assistance: Users can receive empathetic guidance to navigate challenging situations.
• Content creation: The model excels in writing, editing, and generating creative ideas.
• Educational support: It enhances tutoring and provides knowledge-based assistance.
• Art & culture: Accurately identifies and explains historical artworks and references.
• Advanced reasoning and planning: Demonstrates superior performance in multi-step issue resolution via AI agents.

Future Implications and AI Safety

GPT-4.5 is part of OpenAI’s ongoing commitment to AI safety and alignment. Utilizing advanced supervision techniques combined with reinforcement learning from human feedback (RLHF), extensive safety testing was conducted prior to deployment. Future iterations will continue to merge enhanced reasoning with profound knowledge, propelling AI intelligence toward new horizons.

How to Use GPT-4.5

For those eager to experience GPT-4.5:

• In ChatGPT: Currently available to Pro users, it will soon roll out to Plus, Team, Enterprise, and Edu users over the next few weeks. It supports search, file and image uploads, and even code writing in Canvas; however, multimodal features such as voice or video are not yet implemented.
• In the API: Accessible in Chat Completions, Assistants, and Batch APIs, GPT-4.5 supports function calling, structured outputs, streaming, and systematic messaging. Developers can harness its strengths for creating emotionally intelligent responses and facilitating multi-step planning.

Conclusion: What’s Next for AI?

GPT-4.5 stands as a milestone in the journey of scaling unsupervised learning and enabling intuitive AI interactions. OpenAI is closely monitoring user feedback to evaluate long-term API availability based on cost considerations and utility. This model not only pushes the boundaries of AI intelligence but also aims to deliver deeper, more human-like interactions.

I invite all developers, AI enthusiasts, and business leaders to explore the wonders of GPT-4.5, share their feedback, and remain engaged with the transformative role AI continues to play in technology and society.