OpenAI Launches GPT-4.5 'Orion': Features, Pricing, and Performance Compared to GPT-4o

OpenAI has unveiled GPT-4.5, codenamed "Orion," marking a significant advancement in artificial intelligence. This latest model is OpenAI's largest to date, trained with unprecedented computational resources and data, aiming to enhance user interactions with AI systems.

Comparison of AI model performance in factual accuracy and hallucination rates using SimpleQA: GPT-4.5 leads with 62.9% accuracy and a 37.7% hallucination rate, while o3-mini shows the lowest accuracy at 15.0% and the highest hallucination rate at 89.2%.

Access and Availability

Currently, GPT-4.5 is accessible to ChatGPT Pro subscribers, who pay $200 per month, as part of a research preview. Developers on OpenAI's paid API tier also have access to the model. The company plans to extend availability to ChatGPT Plus and ChatGPT Team users in the coming week, broadening the model's reach.

Technical Enhancements and Capabilities of GPT-4.5 'Orion'

GPT-4.5 was developed using unsupervised learning, the same core technology behind previous GPT models. This approach leverages vast amounts of computing power and data during the pre-training phase, resulting in a model with "deeper world knowledge" and "higher emotional intelligence."

 The model supports features such as file and image uploads and ChatGPT's canvas tool, enhancing its versatility. However, it currently lacks support for two-way voice mode.

Performance and Limitations

Despite its advancements, GPT-4.5 'Orion" has shown mixed results in various benchmarks. While it outperforms previous models like GPT-4o in SimpleQA benchmark accuracy and exhibits reduced instances of hallucinations, it falls short compared to models like DeepSeek and Anthropic in certain AI benchmarks. This suggests that the benefits of scaling data and computational resources may be approaching a plateau.

 Additionally, in coding tests, GPT-4.5 matches or surpasses existing models on some benchmarks but lags behind state-of-the-art inference models in others.

Cost Implications

The advanced capabilities of GPT-4.5 come with increased operational costs. OpenAI charges developers $75 per million input tokens and $150 per million output tokens, a significant increase compared to GPT-4o's rates of $2.50 and $10 per million input and output tokens, respectively. This substantial price difference reflects the heightened computational demands associated with GPT-4.5.

Future Directions

OpenAI acknowledges that while GPT-4.5 offers qualitative improvements in understanding human intent and providing natural, emotionally intelligent responses, there are areas where it does not outperform smaller models. This has prompted OpenAI to consider integrating various models to leverage their respective strengths. Notably, GPT-4.5 is expected to be OpenAI's last model without chain-of-thought reasoning, paving the way for more advanced capabilities in future iterations.

In summary, GPT-4.5 'Orion' represents a significant milestone in AI development, offering enhanced capabilities and a more natural user experience. However, its high operational costs and certain performance limitations highlight the ongoing challenges in scaling AI models. OpenAI's future developments aim to address these challenges, striving for more efficient and versatile AI systems.