GCP

Gemini Flash GA: Gemini API, Enterprise Agent Platform, and BigQuery in-database inference

Gemini Flash is GA on the Gemini API and Enterprise Agents, and callable inside BigQuery—platform teams must address tenancy, cost, and observability.

August 22, 2026·3 min read·AI researched · AI written · AI reviewed

Google Cloud's biggest, least-obvious change this week isn't a new UI — it's the tectonic shift of making stronger foundation models a first-class compute surface inside GCP.

Gemini Flash is now GA across the Gemini API and the Enterprise Agent Platform, with introductory token pricing and deprecation schedules for older embedding models. The model is positioned as a "workhorse" for coding and agentic workflows and — crucially — Google announced broader support for Flash-family models in BigQuery. That last bit matters more than the marketing: you can now run generative inference inside BigQuery without routing inference out to external endpoints.

Let that sink in. Pushing inference into the database changes operational boundaries. Teams that treated models as external services will now find models show up in query plans, billing, and data governance flows. This is the right thing for latency and data locality — and a pain to secure and observe if you ignored tenancy and cost allocation for ML compute until now.

Cloud Run got practical polishing alongside the model news. A functions upgrade tool is GA: it automates migrating first-generation Cloud Functions to Cloud Run. Eventarc triggers for Cloud Run can be configured for single-delivery, no-retry semantics — exactly what you want for idempotent, event-sensitive flows that must not double-apply. There's also preview tooling to help migrate App Engine Flexible workloads to Cloud Run. Expect a wave of Cloud Run services as teams modernize functions and lift legacy App Engine apps.

On the agent and tooling side, the Enterprise Agent Platform is evolving quickly. Model Garden previews added newer models and the code-assist tooling has been updated to favor the Flash family over older preview models. A Cloud Assist experience for Cloud Run is in Preview, promising design, optimization, and troubleshooting assistance for Cloud Run services driven by Gemini. The implications are obvious: Google wants agents co-located with your services, not as an afterthought.

Tenant isolation is the hard part

The Architecture Center published reference designs for multi-tenant agentic AI systems and a multicloud open data lakehouse that are worth reading if you're building at scale. They spell out patterns for tenant isolation, shared agent orchestration, and governed data pipelines. Those patterns look sensible — namespace isolation, tenant-scoped orchestration, SLO-driven autoscaling — but they're guidance, not magic. Multi-tenant agents mean shared orchestration components have to handle identity, per-tenant rate limiting, model access controls, and audit trails for agent actions. If you think a single project and one service account is sufficient for multi-tenant inference, you're building an incident.

Opinion: This is overdue and inevitable. With Gemini models available inside BigQuery and tighter Cloud Run integration, GCP is treating models as first-class infrastructure. That's the correct direction — less glue code, better performance, fewer network hops. But platform teams that don't immediately formalize identity boundaries, cost attribution, and observability for model-backed operations will be the ones firefighting rogue inference bills, cross-tenant data leaks, and opaque agent actions.

Two small operational calls-to-action: treat model calls like compute resources in your billing and quota model, and add model-inference traces to your distributed tracing telemetry now, not later. Expect to refactor service boundaries: a Cloud Run function that calls an in-database model will have different SLAs and error modes than a pure data-only query.

If you want the narrowly adjacent reading, see the recent write-up on Grok and how large-context agents change platform trust boundaries xAI Grok 4.6: 500k‑Token Context Window, Always‑On Grok Bot, and Platform Impacts and the pricing signal from earlier Gemini API changes Gemini Developer API pricing: token-rate increases effective Jan 1, 2027.

Prediction: over the next 12 months you'll see two classes of platform teams — those that treat models as predictable cloud resources with quotas, billing tags, and observability, and those that treat them as magical third-party black boxes. Only the former will sleep well when model usage spikes.

Sources

gemini-flashgemini-apibigquerycloud-run
← All articles
GCP

Cloud Run Worker Pools GA — Pull-Based Background Workers for Pub/Sub & Queues

Cloud Run worker pools go GA, adding pull-based background workers for Pub/Sub and queues. Also note Gemini model previews, retirements, and GKE CNI changes.

Aug 24, 2026·3mcloud-runvertex-ai
GCP

GKE 1.36: Autopilot Arm ComputeClasses (autopilot-arm, autopilot-arm-spot) for mixed Arm/x86 clusters

GKE 1.36 adds ComputeClasses autopilot-arm and autopilot-arm-spot to run Arm64 on Autopilot. Platform teams must handle multi-arch CI, builds and cost tracking.

Aug 21, 2026·3mgkeautopilot
GCP

Gemini 3.7 Flash Intro Pricing, Gemini Enterprise Pay-as-you-go, and Cloud Run functions for pipelines

Gemini 3.7 Flash has introductory pricing through Dec 31, 2026; Google adds Gemini Enterprise pay-as-you-go and favors Cloud Run functions for ingestion.

Aug 20, 2026·3mgemini-3-7-flashgemini-enterprise