Gemini Interactions API Is Now GA: The Flex vs Priority Tier Routing Decision Every Operator Must Make

The Gemini Interactions API has reached general availability and is now Google's primary interface for Gemini models and agents. That is not just a stability milestone — it introduces two infrastructure changes that are directly routing-relevant: a Flex tier that cuts inference cost by 50% for background workloads, and managed server-side session state with 55-day retention on the paid tier.

Teams routing Gemini traffic through an AI gateway need to map their workloads to these tiers before they matter.

What happened

Google announced GA for the Interactions API, which launched in public beta in December 2025. With GA, the API is now the default entry point for Google AI Studio and all Gemini API documentation. The older Generate Content and Chat Completions endpoints remain available during a migration window, but new development is directed at the Interactions surface.

The GA release added several capabilities that developers requested during beta:

• Flex and Priority tiers — Flex runs inference in available background capacity at 50% of standard pricing; Priority tier guarantees reserved capacity with lower latency variance.
• 55-day interaction history retention on the paid tier, making long-running conversational sessions retrievable without client-side storage.
• Background execution — interactions can continue processing after the initial response returns, with results retrievable via interaction ID.
• Multimodal generation in a single surface: image generation with Nano Banana 2 and Image Search grounding, music generation with Lyria 3, and multi-speaker TTS.
• Error pinpointing — errors now reference the exact field that caused them, instead of a generic validation failure.

The API schema was significantly revised in May 2026 ("From Roles to Steps"): every action — user input, tool call, model output, thought — is now a typed step rather than a role. Teams that built against the beta schema need to complete that migration before the old schema endpoints close.

Why it matters for AI engineering teams

The Flex/Priority split is a routing policy question. Google's Interactions API now has two cost-performance tiers with meaningful separation. Flex at 50% pricing is designed for workloads where the caller can tolerate variable latency — batch evaluation runs, scheduled report generation, background agent tasks, document indexing pipelines. Priority tier is for user-facing calls where consistent latency matters: interactive chat, streaming completions to an active session, real-time tool calls.

This is the same pattern AWS Bedrock introduced with on-demand vs provisioned throughput, and that Anthropic explored with fast mode — but the 50% pricing delta is larger than most provider tier splits to date. For teams running significant Gemini volume, the workload classification exercise is now a cost-reduction opportunity, not an optional optimization.

Managed state changes what you need to store. The 55-day paid-tier retention means interaction history is retrievable from Google's infrastructure without the caller maintaining a session database. For teams currently storing Gemini conversation histories in their own databases or key-value stores for context reconstruction, this is worth evaluating: offloading that storage to the provider simplifies the client-side implementation but creates a provider dependency for session continuity.

Background execution affects how you structure agent pipelines. When an Interactions API call triggers a long-running chain — a deep research task, a multi-tool orchestration sequence — the response can return immediately with an interaction ID, and the result is available for retrieval when processing completes. For agent pipelines that previously blocked on Gemini completion, this changes the polling/callback architecture.

The May 2026 schema change is still active. Teams that built against the Interactions API beta before May 2026 used the roles-based schema (user, model, function). The GA schema uses typed steps. If you have beta Interactions API integrations in production, confirm they are on the GA schema before the old endpoints close.

The router/operator angle

Tier routing policy: If you route Gemini requests through an AI gateway or proxy, the Flex vs Priority selection is a per-request parameter — not a global account setting. Your routing layer must pass the appropriate tier to the Interactions API based on workload classification. A request tagged as background or batch should route to Flex; a user-facing streaming call should route to Priority.

Provider fallback on Flex: The Flex tier uses available Google capacity, which means throughput is not guaranteed at peak times. Teams that use Gemini for latency-tolerant workloads should retain a fallback model (e.g., a self-hosted or alternative provider API) for when Flex capacity is saturated and the timeout is reached.

Multi-modal routing consolidation: The GA announcement adds image, music, and speech generation to the same Interactions API surface. For teams currently routing image and audio requests through separate Google endpoints (Imagen, Lyria, TTS), consolidating to the Interactions API is now the intended path. Verify whether your gateway's model-routing table needs updating to map those modalities to the new endpoint.

Internal link check: The Interactions API docs live at /docs/ on ai.google.dev and are not hosted on therouter.ai. Link to the Google source directly rather than an internal path.

What TheRouter users should watch

Teams routing Gemini workloads through TheRouter that have not yet classified traffic by latency sensitivity are leaving 50% cost reduction on the table for non-interactive calls. The Flex/Priority tier split is the most direct lever available today for reducing Gemini inference costs without changing models.

For teams evaluating whether to migrate conversation storage from client-side databases to Interactions API managed state: the 55-day retention covers most active user windows but not long-term archival — plan accordingly if your use case requires retrieval beyond two months.

If you use Gemini for multi-modal generation (image, audio), confirm your SDK and gateway configuration now references the Interactions API rather than the legacy generation endpoints, as those endpoints will be deprecated on a timeline to be announced.