A Source-Checked Analysis of Google's Gemini Models on Long Context, Grounding, Cost, and Device Fit

Google's Gemini model family has grown from multimodal assistants into cloud image and video generators, on-device models and tiered paid plans. This preprint is a source-checked analysis of public claims about these models. It compares vendor specifications with independent evidence in five areas. First, long-context retrieval, where single-needle recall results contrast with lower reported multi-needle scores. Second, search grounding, where studies of Google's AI Overviews find that accurate answers are often not fully supported by their citations; one analysis reports about 91% accuracy but only 39% of overviews both correct and fully supported. Third, cost per task rather than cost per token. Fourth, naming, plan and pricing changes that make it hard to know which model is in use. Fifth, device fit across Gemini Nano, Flash, Pro, Argon and the Nano Banana image models. Primary sources, including Google documentation, technical reports and an original study, are separated from secondary reports, and unverified claims are flagged. The paper ends with a practical model-selection guide and four open questions for developers and Google. It is a non-peer-reviewed analysis of public information as of November 5, 2026, and it reports no new experiments.