The third week of April 2025 brought a new family of OpenAI models, a rewritten safety framework from the same company, a Google model trained on dolphin vocalizations, and another public estimate of when human-level AI might arrive. Here is a summary of the verified developments from April 14 to 20.
OpenAI releases the GPT-4.1 family
On April 14, OpenAI introduced GPT-4.1, GPT-4.1 mini and GPT-4.1 nano, three models available only through its API. The company said the models improve on GPT-4o in coding, instruction following and long-context comprehension, and support a context window of up to one million tokens. OpenAI said GPT-4.1 scored 54.6 percent on SWE-bench Verified, a software engineering benchmark, compared with 33.2 percent for GPT-4o.
The smaller mini and nano versions are priced for high-volume, lower-latency tasks, with nano described by OpenAI as its fastest and cheapest model to date. At the same time, the company said it would retire the original GPT-4 model from ChatGPT on April 30 and phase out the GPT-4.5 Preview in the API by July, since GPT-4.1 offered similar or better performance at lower cost.
A revised Preparedness Framework
On April 15, OpenAI published an updated version of its Preparedness Framework, the internal process it uses to track and mitigate severe risks from frontier models. The new version narrows the focus to capabilities that could cause severe harm, such as biological and chemical misuse, cybersecurity and AI self-improvement, and sets two thresholds: “high” capability, which requires safeguards before deployment, and “critical” capability, which requires safeguards during development as well.
The update also adds research categories for behaviors the company says it is monitoring but does not yet treat as tracked risks, including models that hide their capabilities, evade safety measures or resist shutdown. OpenAI said it would publish safeguard reports alongside its capability reports. Critics noted that the revised framework also says the company may adjust its requirements if a competitor releases a high-risk system without comparable protections, a clause that drew scrutiny from safety researchers.
Google’s DolphinGemma
On April 14, Google announced DolphinGemma, a roughly 400 million parameter audio model built on its Gemma architecture and trained on decades of recordings from the Wild Dolphin Project, a long-running study of Atlantic spotted dolphins in the Bahamas. Developed with researchers at Georgia Tech, the model is designed to identify structure in dolphin vocalizations and predict the sounds likely to follow, in the same way a language model predicts the next word.
Google said the model is small enough to run on Pixel phones that researchers use in the field, and that it planned to release it as an open model in the summer. The company described the work as a step toward understanding dolphin communication, not a translation device, and said whether dolphins have anything resembling language remains an open scientific question.
Hassabis on AGI, and Alphabet’s spending
Demis Hassabis, chief executive of Google DeepMind, repeated in interviews during the week, including one broadcast by CBS’s 60 Minutes on April 20, that he expects artificial general intelligence, a system with the full range of human cognitive abilities, within five to ten years. He paired the forecast with a call for international coordination on safety and said society is not yet prepared for the change.
Alphabet, Google’s parent, has committed to spending about $75 billion on capital expenditures in 2025, mostly on AI data centers and chips, a figure chief executive Sundar Pichai first disclosed in February and reaffirmed at the company’s Cloud Next conference in early April despite investor concern about the scale of the outlay.
Taken together, the week showed the two tracks the industry is running on at once: faster, cheaper and longer-context models for developers, alongside more formal attempts to define what risks those models might pose and how to measure them.
