Launch of GPT-4.5

OpenAI announced the launch of GPT-4.5, the largest AI model ever created by the company, code-named Orion. This new model has been trained using more computing power and data than previous models. However, OpenAI does not consider GPT-4.5 a frontier model. ChatGPT Pro subscribers can access GPT-4.5 starting today as part of a research preview, while developers on paid tiers of OpenAI’s API can use it immediately. ChatGPT Plus and ChatGPT Team users will gain access to the model next week. The industry is eagerly awaiting Orion, seen as an indicator of the validity of traditional AI training approaches. GPT-4.5 was developed using similar techniques to those used for earlier models, dramatically increasing computing power and data during the pre-training phase. Previous GPT generations showed significant performance improvements across various domains, such as mathematics and writing. OpenAI claims that GPT-4.5 has a deeper world knowledge and higher emotional intelligence, although progress seems to be slowing compared to newer reasoning models.

GPT-4.5 has very high operational costs, prompting OpenAI to evaluate the long-term sustainability of offering the model. The company charges $75 per million input tokens and $150 per million output tokens, compared to much lower costs for GPT-4o. OpenAI clarifies that GPT-4.5 is not a direct replacement for GPT-4o, the main model for most of its applications. While it supports features like file and image uploads, it currently lacks support for two-way voice mode. However, GPT-4.5 has shown to be more performant than GPT-4o on benchmarks like SimpleQA, where it achieved better accuracy results. OpenAI states that GPT-4.5 hallucinates less frequently than other models, which suggests a lower likelihood of providing incorrect information.

Despite these successes, GPT-4.5 does not reach the performance of some more advanced reasoning models on challenging academic benchmarks. However, it performs well on math and science problems, showing to be qualitatively superior in areas that benchmarks do not capture, such as understanding human intent. In informal tests, GPT-4.5 demonstrated greater emotional intelligence, responding more appropriately to sensitive situations. OpenAI acknowledges that academic benchmarks do not always reflect real-world usefulness and is committed to further exploring GPT-4.5’s capabilities. Challenges related to scaling laws suggest that pre-training may not continue to be effective as in the past, leading the industry to consider more complex reasoning models.