Google

Google's current platform comparison names the Gemini Developer API and Gemini Enterprise Agent Platform API. Both use the Google GenAI SDK. Cloud deployment adds service-account authentication and different region and enterprise-control considerations. Older integrations may still say Vertex AI; the generateContent cloud example retains vertexai=True. Enterprise-platform Interactions is separately in Preview and currently global-only. The naming change alone is not an instruction to rewrite a working client.

The Interactions API is generally available since June 2026 and Google recommends it for new projects. It can continue stored conversation history through previous_interaction_id. That history includes inputs and outputs; tools, system instructions and generation settings must be supplied again for each interaction. An application that continues an investigation therefore still needs to provide the instructions and tools intended for the new turn.

The older generateContent API remains supported and has a specific continuing use. Google's caching documentation says Interactions supports automatic, implicit cache reuse, while creating and managing explicit cache objects requires generateContent. Repeated input can qualify for implicit caching, but a similar prompt does not guarantee a hit. The response's cached-token usage field shows what was reused, making it more useful than assuming a saving from the request alone.

Gemma is a separate open-model family whose weights can run on a team's hardware or a hosted service. It shares research lineage with Gemini. Choosing Gemma changes the deployment question toward the runtime and hardware serving those weights; it does not give that deployment the Gemini app's entitlements.

The consumer Gemini app has separate activity and privacy controls. With Keep Activity enabled, activity can be used for model training. Turning it off excludes future chat training unless feedback is submitted, but chats remain stored for 72 hours and can be reviewed for protection purposes. Temporary chats also remain for 72 hours and are not used to train Google's AI models. Work and school accounts may have different terms. These controls are distinct from the API's paid-service policy.

The Gemini API terms distinguish unpaid and paid services. Unpaid input and output can be used for product and machine-learning improvement and reviewed by people. Paid prompts and responses are not used to improve Google products, but limited safety and legal logging still applies. API paid status depends on an active billing-linked project; AI Studio can fall under paid terms through access to such a project or a Workspace enterprise account. For users in the EEA, United Kingdom and Switzerland, the paid data-use terms also apply to services offered free of charge. The account and region matter alongside the price displayed.

Google Cloud's enterprise-platform data guidance says managed-model data is not used for training or fine-tuning without prior permission or instruction. It also documents feature-specific retention: abuse monitoring, advanced-model safeguards, grounding and optional request-response logging have separate conditions. For example, Google Search grounding retains relevant logs for up to three days with no storage-disable option for that feature. A training restriction does not by itself establish zero retention.

Gemini Notebook has its own privacy notice. It says notebook content is not used to directly train foundational models unless feedback is provided. Feedback can include sources, uploads and outputs, with human review and retention for up to three years. Workspace and Education users have a separate protection against human review and model training, even for feedback. Data shared to another Google service follows that service's notice.

Interactions also has a storage choice. Its retention documentation defaults to stored requests: 55 days on the paid tier and one day on the free tier, with shorter paid retention options. Setting store=false opts out of interaction storage, but prevents later continuation through the previous ID and is incompatible with background execution. That trades server-side continuity for stateless requests.

Gemini API rate limits apply per project, not per API key. The usual dimensions are requests per minute, input tokens per minute and requests per day; daily quotas reset at midnight Pacific time. Models and usage tiers have different ceilings, with additional restrictions for some previews and experimental models. Creating another key in the same project does not create another quota. AI Studio shows the project's active limits, and the documented ceilings do not guarantee available capacity.

Billing adds a different shared limit. Google's billing guide ties keys to their project and aggregates usage across projects under the billing account. Exhausting a prepay balance can stop all linked projects with HTTP 402. Project spend caps are experimental, and billing processing can lag by around ten minutes; batch completions and agent sessions can incur overages. A cap or balance is therefore a spending control with documented delay, rather than a precise per-request financial cutoff.

For a transient 429 or 503, the troubleshooting guide recommends exponential backoff, jitter and a maximum number of retries. It also describes automatic retry behavior in official SDKs. Invalid requests, depleted credits and denied authentication need a different response: repeatedly submitting a 400, 402 or 403 will not fix the cause. A caller can distinguish a temporary capacity problem from a request or billing problem before deciding whether to retry.