Google just pushed Gemini 3.1 Pro into preview across Vertex AI and Gemini Enterprise — and unlike a quiet model bump, they exposed it through the Gemini API in a surprising number of developer touchpoints: Google AI Studio, Android Studio, the Antigravity experimental surfaces, and the Gemini CLI. That means teams won't just see the model in the AI console; developers will be able to call the same preview model from IDEs and local tooling without a separate prod gate.
This is significant for two reasons. First, the policy and telemetry boundary for "who can use what model" has shifted: previews are now reachable from dev tooling that sits on developers' laptops and CI systems. If your org treats preview models as harmless for experimentation, that assumption will be tested — data exfiltration, PII in prompts, and uncontrolled inference costs become operational problems at scale. Second, Gemini Enterprise's preview capabilities around data stores and actions widen what a model can do inside a GCP project, so access is functionally equivalent to adding a new service account with elevated privileges.
Google also surfaced Gemini Enterprise preview updates for additional data stores and actions in the release notes. Details remain sparse at preview level, but the pattern is clear: more persistent state and richer control hooks for models. Treat "actions" as another vector for privilege escalation — they can call APIs and mutate resources if you let them.
Billing changes landed at the same time and will matter to platform teams: Google said resource-based Committed Use Discount (CUD) sharing will default to enabled for new billing accounts and for eligible existing billing accounts without active CUDs, with rollout beginning in mid‑2026. The practical effect is that committed discounts are more likely to be applied across linked projects by default, reducing wasted committed spend but also collapsing cost isolation. If your org uses billing-account separation for strict chargeback, expect reconciliations to need adjustment and to show lower billed costs for accounts that previously absorbed their own committed spend.
This default flip is overdue from a cost-optimization perspective — most large orgs underutilize CUDs because discounts are siloed — but it's careless to enable it without forcing teams to opt out explicitly when they rely on accounting boundaries. Platform engineering should treat the change as a policy event: update billing policies, check IAM on billing accounts, and validate your cost-allocation tags and exports.
Other release-note noise worth noticing: Cloud Service Mesh had a recent maintenance update, and the Cluster Toolkit added integrated cluster health checks, static node-count collection for blueprints, and stricter GKE node auto-provisioning accelerator validation. Those Cluster Toolkit changes matter if you run automated cluster provisioning pipelines; they tighten validation paths that previously allowed misconfigured accelerators. If you haven't already, read the toolkit update and consider adding those health checks into your provisioning pipeline: GKE Cluster Toolkit adds integrated health checks and stricter NAP accelerator validation.
What to do this week
- Inventory the surfaces where Gemini 3.1 Pro preview is reachable: project APIs, OAuth clients, and developer extensions (Android Studio, CLI installs). Lock down who can call the Gemini API and require audit logging for preview endpoints.
- Treat Gemini Enterprise "actions" the same way you treat service account permissions: map actions to effective API calls and test for privilege boundaries.
- Update billing-account policies and dashboards for the CUD-sharing default. If you rely on per-account committed-spend accounting, prepare an opt-out and adjust FinOps processes.
Opinion: Google is right to expand where the model is usable — developer ergonomics matter — but they're pushing a production-grade model into developer surfaces without forcing org-level guardrails. That accelerates time-to-feedback, but it also hands platform teams a new emergency category: "model-led incidents" that combine security, cost, and compliance. If you aren't treating model access like a service endpoint with RBAC, audit logs, and billing alerts, your next on-call pager will be a model misuse incident.
If your team keeps a tight leash on production models, this preview will be a boon. If you don't, expect a month of surprises and a renewed demand for org-level controls that stop being optional. Google has given you power — now make it governable.