Vertex AI

Google's AP2 sample repository splits its authentication guidance in two: a Google API key for development, Vertex AI for production deployments. That split says most of what Vertex AI is in practice, the GCP surface where picking a model stops being a convenience and becomes an operational commitment.

The model announcements follow the same pattern. When Google made Gemini 2.5 Flash-Lite stable and generally available in July 2025, priced at $0.10 per million input tokens and $0.40 per million output tokens, it shipped through Google AI Studio and Vertex AI. Multi-model selection is the platform's stated function: choose the model the workload calls for, from low-latency classification and translation up to deeper reasoning.

The harder question sits above that selection. Janakiram MSV reads the enterprise agent stack as three vendors converging on one architecture, with Gemini Enterprise Agent Platform alongside Bedrock AgentCore and Microsoft Foundry each assembling runtime, memory, tool gateway, identity, observability and governance under different names. No portability contract joins them. Session state, traces and identity all land with a single provider, and moving an agent a year later means rebuilding it. Mathieu Grymonprez, Global CDO of Adeo, names that lock-in of enterprise intelligence as his biggest worry, while granting Google a "conscience de la prod" he does not yet find at OpenAI or Anthropic.

Open protocols such as MCP supply many of the primitives. The lifecycle above them, versioning, promotion, rollback, is still whoever's platform you deployed on.