Ad
Short

OpenAI is planning to release three new AI models: GPT-4.1, o3, and o4 mini, according to a report from The Verge citing sources familiar with the company’s roadmap. GPT-4.1 is described as an upgraded version of GPT-4o, the multimodal model introduced last year that processes text, audio, and images in real time. In addition to the full model, OpenAI is expected to offer two smaller variants: GPT-4.1 mini and GPT-4.1 nano. The company is also preparing to launch the full version of the o3 reasoning model, along with a compact variant called o4 mini, which might be released first. References to these models have already been discovered in an updated web version of ChatGPT.

Ad
Ad
Short

OpenAI plans to remove its GPT-4 language model from ChatGPT on April 30, according to a recent changelog announcement. The company will replace it entirely with GPT-4o, which OpenAI says performs better than the original model across writing, programming, and STEM tasks. While GPT-4 will disappear from ChatGPT, developers can still access it through OpenAI's API. The 2023 model, which CEO Sam Altman indicated cost over $100 million to develop, remains at the center of several ongoing copyright disputes, including a lawsuit from the New York Times.

Ad
Ad
Ad
Ad
Short

OpenAI has introduced a new initiative called the "Pioneers Program" aimed at developing AI benchmarks tailored to specific industries. The company says the goal is to create evaluation methods that better reflect real-world use cases in areas such as law, finance, and healthcare—domains where existing benchmarks fall short. According to OpenAI, current AI benchmarks are often flawed. They tend to measure tasks that are difficult to interpret or overly susceptible to manipulation—criticisms that have also been directed at OpenAI itself. As reported previously, the company has faced scrutiny over its involvement in funding and promoting a prominent math evaluation dataset. In the coming months, OpenAI plans to collaborate with multiple companies to build domain-specific evaluation tools. These benchmarks will eventually be released publicly. The first cohort includes select startups focused on practical AI applications. Participating companies will also have the opportunity to work with OpenAI on improving model performance via reinforcement fine-tuning, a method the company recently introduced for customizing expert-level language models.

Google News