Google has published a comprehensive guide to the lifecycle and upgrade paths of generative models—including the Gemini and embedding model families—on Vertex AI. The documentation introduces essential terminology, clarifying distinctions between stable models, legacy stable models, and retired models. Stable models are designated as publicly released and fully supported for production, with a clear retirement date announced for each. The ´latest stable´ model within a family is recommended for new and active projects, whereas ´legacy stable´ models, having been superseded, remain available but may face access restrictions for new projects. Retired models are no longer accessible via API after their official retirement date.
The guide identifies the latest stable models, including major Gemini releases such as gemini-2.5-pro and gemini-2.5-flash (available from June 17, 2025), alongside other embedding options like gemini-embedding-001 and text-embedding-005, all listed with release and retirement timelines. In contrast, legacy stable models—such as gemini-1.5-pro-002 and its flash variant—carry suggested upgrade paths. Migration is crucial as retirement dates approach, and Google recommends two approaches: a structured migration process, prioritizing risk minimization and performance validation, and a quick migration method for urgent cases, which swaps model IDs and verifies core functions with minimal delay but higher potential risk.
The document also explains the use of auto-updated aliases, which allow applications to always point to the latest stable Gemini model without requiring code changes upon new releases. Several alias mappings are provided as current references, ensuring continued compatibility as the platform evolves. For transparency, an expandable section lists all retired models, their life spans, and the corresponding recommended upgrades. Developers are urged to migrate off retired models well before access is revoked to avoid interruptions. Additional resources direct users to model availability by region and detailed documentation on individual model capabilities and endpoints, reinforcing Google’s commitment to transparent lifecycle management in artificial intelligence infrastructure on Vertex AI.
