Google's new experimental AI model for advanced math problems

A leaked configuration has revealed Google's internal experimental AI model, codenamed Mathematica, which is based on the DeepThink V3 architecture. The model is designed to handle complex symbolic computations and heavy mathematical problem-solving. It supports a context window of up to 1 million tokens and has an output limit of 65,536 tokens. The model is marked as 'UNSTABLE_EXPERIMENTAL,' indicating it is an unstable experimental version. It also carries the 'Teamfood' tag, used internally for employee testing. The leak comes after Google launched the Gemini DeepThink IMO mode for high-level math competitions and released Gemini 3.8 Live and Live Extended Thinking on September 15, 2026. According to API data, the model's internal identifier is 'models/deepthink-mathematica-tf-raw-thoughts.' IT Home notes that the context window refers to the total amount of text the model can process in one go, while the output limit is the maximum text it can generate in a single response. The leak was shared by user @lyraxana on September 16, 2026, on the X platform.