Cloud Run just stopped being just a scale‑to‑zero request platform. The new "instances" preview gives you long‑lived, individually addressable units with explicit per‑instance cost examples (an instance with 1 vCPU and 1 GiB continuously for 30 days ≈ $5.70). That number is small enough to change architectural tradeoffs: always‑on agents, persistent socket connections, and background workers now belong in "serverless" alongside bursty request handlers.
What Google shipped — and why it matters
Instances in Preview: addressability + steady cost
The feature exposes Cloud Run instances you can target directly (not just via stateless request invocations) and documents a per‑instance cost profile. Conceptually this blurs the line between ephemeral container invocations and small VMs: you keep the operational simplicity of Cloud Run (runtime, autoscaling controls, platform managed) while accepting a modest steady charge per instance.
Operational implications are immediate. If your team has been stretching Cloud Run to host always‑on agents (background workers, websockets, or long‑running connections) you can now model those as cheap single‑vCPU instances and stop inventing hacks (cron hacks, external keep‑alives, or opaque warm pools). If you run dozens of agents, recompute — the per‑instance steady cost scales. If you run hundreds, it doesn't take long to prefer a different execution plane.
Gemini and Vertex AI agentization
At the same time Google is widening Gemini and Vertex AI availability. Recent Gemini Flash updates add multimodal and agent-friendly capabilities exposed via the public Generative/Interactions APIs. Put together with Cloud Run instances (and previews that integrate Cloud Run with Gemini tooling), the pattern is clear: always‑on small compute hosting an agent instance that calls out to a multimodal LLM when needed.
Google is also steering users toward a centralized Agent Platform and managed agent APIs as the canonical way to run and govern these agents. That's sensible — centralizing lifecycle, policy, and observability avoids a million ad‑hoc agent control planes — but it's also a new critical dependency and control plane to harden. Treat it like a database: design for failure, auditing, and performance limits.
GKE networking changes — the firewall wrinkle
The release notes also include platform changes to GKE: a recent minor release modified how automatically created firewall rules behave for Services. That sounds minor until you realize many teams assume Kubernetes service exposure is only a Kubernetes concern. Automatic firewall behavior changing means cluster operators need to revalidate their service‑to‑firewall mappings, especially for hybrid connectivity, private clusters, and external load balancers.
If your infra still relies on implicitly created rules, audit now. Fix your IaC to manage firewall rules explicitly, or lock down Service exposure via NetworkPolicy and private control planes. This will bite teams that upgraded without a preflight check.
Securing inference on GKE and storage guidance
Google published guidance on securing AI inference on GKE, recommending a model‑protection choke point, authenticated proxies, enforced egress controls, and sizing storage for the I/O patterns of large model loads. If you mix Cloud Run instances for orchestration and GKE for heavy inference, expect to design a secure, high‑throughput bridge between the two.
What to do this week
- Recompute costs for any always‑on Cloud Run use: treat an instance as a cheap VM and model at scale.
- Audit GKE service exposure and firewall rule assumptions before upgrading to a newer cluster release.
- If you plan agentic deployments, adopt the centralized Agent Platform and treat managed agent APIs as dependencies that need redundancy and RBAC controls.
- Apply a model‑protection layer, enforce network egress controls for GKE‑hosted inference, and optimize storage for model load patterns.
This is the right push from Google: making agent deployments first‑class and integrating generative models with simple serverless runtimes removes a lot of brittle engineering. But it also centralizes a new control plane and mixes VM‑style billing into serverless economics — platform teams that pretend nothing changed will be surprised by cost and network fallout. Expect the next 12 months to be dominated by teams leaning into small, always‑on serverless agents or carving them out into purpose‑built fleets; the middle ground (opaque warm pools and credential ad‑hoc hacks) will evaporate.
If you want the early reading on the Gemini side of this, see our piece on recent Gemini Flash updates and agentic video features Gemini Flash and Agentic Video Features.