Keep pulling the thread on Mark Chen.
The performance improvement of GPT-4.5 over previous models is commensurate with the jump from GPT-3.5 to GPT-4.
OpenAI is scaling its models along two different axes: unsupervised learning and reasoning.
Over the last 1.5 to 2 years, OpenAI has found a new paradigm through reasoning, which it is also scaling.
GPT-5 could be the culmination of OpenAI's work in both unsupervised learning and reasoning coming together.
GPT-4.5 is proof that the scaling paradigm for unsupervised learning can continue by adding more compute, algorithmic efficiencies, and data.
In user preference tests, GPT-4.5 is preferred over GPT-4.0 by a 60% margin for everyday use cases.
For productivity and knowledge work, GPT-4.5 has a 70% preference rate against GPT-4.0.
OpenAI is not seeing diminishing returns from adding more compute and data to its models.
The inference costs for GPT-4 have dropped by many orders of magnitude since its initial launch.
OpenAI is releasing its new model, GPT-4.5, today.
GPT-4.5 is OpenAI's largest and best model to date.
GPT-4.5 represents an order of magnitude improvement over previous models, commensurate with the jump from GPT-3.5 to GPT-4.