The Allen Institute for AI (Ai2) has launched Olmo 3.1, an advanced iteration of its Olmo model family, which enhances reinforcement learning training to improve reasoning benchmarks. This update includes two optimized versions, Olmo 3.1 Think 32B for advanced research and Olmo 3.1 Instruct 32B for instruction-following tasks, alongside a programming-focused model, Olmo 3-Base.

Researchers at Google and UC Santa Barbara have developed a new framework that enhances the efficiency of AI agents in utilizing their compute and tool budgets. This framework includes techniques such as the 'Budget Tracker' and 'Budget Aware Test-time Scaling,' which help agents manage their reasoning and tool-use allowances effectively.

OpenAI has released GPT-5.2, showcasing significant advancements in deep reasoning and coding capabilities, while receiving mixed reactions from early testers regarding its conversational abilities. The update aims to enhance productivity for business tasks and workflows, particularly in coding and data management.

OpenAI has officially launched GPT-5.2, its latest large language model, which is touted as the most capable model series for professional knowledge work. This release comes in response to Google's Gemini 3, which recently gained significant traction, amassing 200 million users within three months of its launch. OpenAI emphasized that the timing of GPT-5.2's release was planned well in advance of Gemini 3's debut.

Marble, a startup focused on developing artificial intelligence agents for tax professionals, has secured $9 million in seed funding to address the labor shortage and regulatory challenges in the accounting industry. The funding round was led by Susa Ventures, with contributions from MXV Capital and Konrad Capital.

Nous Research has launched Nomos 1, an open-source AI mathematical reasoning system that achieved a score of 87 points on the William Lowell Putnam Mathematical Competition, ranking second among 3,988 participants. This competition is renowned for its difficulty, with the top score being 90 and the median score just 2.

Cohere has launched Rerank 4, an advanced version of its search model, which features a context window of 32K, quadrupling the capacity of its predecessor, Rerank 3.5. This enhancement allows the model to process longer documents and evaluate multiple passages at once, improving the accuracy of information retrieval for agents.