Google just slipped a couple of small, high-impact changes into their release notes and — as usual — they matter most where automation assumes stable addresses and free telemetry.
The one you need to notice today: model-routing gateways created on or after 2026-09-03 now get Cloud Run–style hostnames in the form https://GATEWAY_ID-PROJECT_NUMBER.REGION.gateway.dev. That sounds trivial until you remember how many systems treat gateway URLs as constants: DNS records, CORS allowlists, client SDKs, service meshes, and synthetic checks. If your delivery pipeline creates gateways on demand, or you replicate stacks across projects/regions, expect fallout unless you update templates and discovery logic.
Why this is nastier than a doc note
This is not a protocol change or API surface innovation — it's a naming convention change with real operational consequences. Teams that hardcode endpoint hostnames, bake hostnames into client configs, or use short-lived gateway resources for blue/green and canary releases are about to see 404s and CORS failures. More pernicious: monitoring and synthetic tests that validate a specific host pattern will quietly fail, while load balancers or WAF rules keyed on old hostnames will stop matching.
The new format is explicitly: https://GATEWAY_ID-PROJECT_NUMBER.REGION.gateway.dev. Take a moment: the project number is now inside the hostname, not the project id. That means simple string substitution scripts based on project ids will break.
App Topology API: billing catch-up
Google says the App Topology API will move to usage-based billing with a daily free data allotment starting 2026-09-15. Topology APIs are inherently chatty when UI dashboards, multi-tenant maps, and automated remediation pull data frequently. Shifting to usage-based billing forces teams to treat topology as a product: cache aggressively, bulk requests, and avoid per-refresh network calls from dozens of clients.
This will hit teams running continuous discovery pipelines and large-scale service-mapping UIs. The free daily allotment softens the blow, but the right architectural response is obvious: snapshot topology at a controlled cadence, push diffs, and minimize synchronous pull patterns from client applications.
GKE defaults and Compute Engine reservation flexibility
Separately, GKE cluster creation defaults have moved in recent releases: several channels now default to Kubernetes 1.35.7, while others remain on earlier 1.35 builds in some regions. Check Google Cloud release notes or your Cloud Console for the exact GKE build numbers in your target channels. Treat the announced versions as the new floor for freshly provisioned clusters; don't assume older behavior when scripting cluster creation or admission logic.
Compute Engine also added the ability to convert single-project reservations to shared reservations and back. That's a practical addition for organizations that flip between tight billing scopes and pooled capacity, and it avoids the previous need to recreate reservations when policy or ownership changes.
What you should do in the next 72 hours
- Audit automation that creates or consumes model-routing gateways for any hardcoded hostname patterns; update templates to derive hostnames dynamically (use project number when needed).
- Check DNS/CNAME automation, WAF rules, and synthetic checks for hostname assumptions and add pattern-matching where appropriate.
- Review App Topology API usage: introduce caching, aggregated calls, or delta endpoints; estimate daily topology volume against the upcoming billing model.
- If you rely on specific GKE behavior, pin cluster versions in IaC when you need predictable behavior; otherwise, accept the new defaults and validate tests against the current 1.35.x builds.
Opinion: small changes, outsized ops risk
Google made technically sensible moves — standardizing gateway hostnames and aligning topology APIs with usage are the right long-term calls. But operations cost lives in the small, brittle assumptions: hostnames, project-number vs project-id, and free telemetry that suddenly costs money. Platform teams that still treat infra as immutable will get burned. Treat this as another reminder: encode discovery not assumptions.
If you run model routing or heavy topology-driven tooling, you should already be updating IaC and caching patterns. If you don't, expect surprises when the next deploy creates a gateway your clients can't reach.
For teams working with Gemini/Vertex AI model routing, this follows other recent Cloud Run and AI changes — see our coverage of Gemini and Cloud Run worker pools for context on how Google is iterating on AI endpoints and serverless containers: Gemini Pro preview in Vertex AI, Flash1Lite rollout, and Cloud Run worker pools GA.
One final thought: naming is part of API stability. When providers change names, they shift the maintenance burden to customers. That's fine — until your incident war room lights up at 2 a.m.