Google Cloud introduced Cloud Run instances on 27 August at 16:00 UTC: "cost-effective, dedicated, singleton compute runtimes designed to support long-lived workloads like personal AI agents." Each instance "runs just one instance with no autoscaling" and gets "a HTTPS URL that remains unchanged across updates and restarts." Google's own documentation uses the third-party open-source agent OpenClaw as the worked example, passing OPENCLAW_GATEWAY_PASSWORD and GEMINI_API_KEY as environment variables.

The number that will travel

"The cost to run a Cloud Run instance with 1 vCPU and 1 GiB of memory continuously for 30 days is $5.70." That is priced to kill the reflex to spin up a small VM, and it is the whole pitch.

What the common framing gets wrong

Four qualifications sit on the announcement page itself. It is preview, not GA — the deploy command is gcloud beta run instances create, and the page closes with "Currently in preview." There is no stated GA date and no SLA. "Dedicated" and the price do not describe the same thing: the page frames the alternative as "running a dedicated VM — means paying for full compute 24/7," then states the $5.70 figure applies to something built on "shared vCPU with vCPU burst budgets." That is a burstable slice being priced against the dedicated compute it is compared with, and the two sentences sit four paragraphs apart. "Long-lived" tops out at seven days — "up to 7-day continuous runtime, with automatic restart policy configured by default." Any agent state not written to the mounted Cloud Storage bucket dies at that restart. This is durable addressing sold as durable runtime. And the only performance figure is a testimonial: "It reduced cold starts by 88%" comes from one customer, OffDeal, with no baseline, no configuration and no methodology. Google publishes no cold-start number of its own.

The missing feature

SSH — the thing that would make this a usable long-lived box — is "coming soon" behind a private-access signup form, for both instances and services.