Skip to content

Google Cloud Adds a Deferred Gemini Tier at Half Price for Background Agents

Google Cloud has made its deferred service tier for Gemini generally available. It targets long-running agent workloads, offering 50% off completed turns and server-side queuing during peak traffic.

3 min readBased on 1 source0 views
CLOUD

TL;DR

  • Google Cloud's deferred service tier on the Gemini platform is generally available, aimed at background AI agents.
  • Google says completed turns cost 50% of Standard pay-as-you-go rates, and cancelled, expired or failed turns are not charged.
  • Requests are queued and retried on the server side, which Google says avoids 429 errors during peak traffic.

The problem with long-running agents

Some agents run multi-hour background loops: they call a model, use a tool, read the result, and call the model again. Google Cloud's weekly update for November 16 to 20 says that running such loops through synchronous, real-time inference adds cost that is not needed. It also says this pattern can trigger 429 resource exhaustion errors at peak times. A 429 is the HTTP status a service returns when a client has sent too many requests or the service has no capacity left for them, so the request is rejected and the caller has to retry or fail.

Google Cloud points to batch inference as the existing cheaper option. Batch offers a 50% discount, but only when every prompt is in a fixed input file before the job starts. An agent loop does not work that way, because each next step depends on the output of the previous one. That gap is what the deferred tier is meant to fill.

How the deferred tier works

The deferred tier gives agent turns batch-style economics while keeping them dynamic, stateful and multi-turn. According to Google Cloud, a scheduler sends latency-tolerant agent turns to spare global capacity, so they do not compete with live chat traffic. A latency-tolerant turn is one that can wait a while for its answer.

To use it, developers set service_tier="deferred" in the RunConfig of the Agent Development Kit (ADK), or pass the same setting to client.interactions.create() in the Interactions API. The post lists gemini-4-argon, gemini-4-carbon, the gemini-3.x models and gemini-deep-research-1.0 as supported. To chain turns and keep prompt caches warm, the developer passes previous_interaction_id.

Pricing and failure behaviour

The company describes three parts of the offer. Completed turns get 50% off Standard pay-as-you-go rates, and this discount stacks with prompt caching, which reuses parts of earlier prompts. Cancelled, expired or failed turns are not billed. During peak traffic, the platform queues and retries turns itself, and Google Cloud says this produces zero 429 errors.

The headline claim of "half the cost" therefore applies to completed turns. Teams should read their own bills against that rule, especially for workloads with many retries or abandoned runs.

What the source does not say

The only source here is Google Cloud's own rolling "What's new" page, which is a short summary with links to guides. It contains no independent benchmark, no measured latency figures, and no information about how long queued turns may wait before they run. The claims about zero 429 errors and spare global capacity are the company's own description and have not been checked against other reports.

Why it matters

For teams that run background agents on Gemini, the deferred tier is a cost option that does not require redesigning the agent loop. The change in code is a single configuration setting, according to the post. The practical question is which of an application's calls can tolerate waiting. Interactive chat should probably stay on synchronous inference. Research jobs, batch-like analysis and background tool loops are the likely candidates to move.

Teams should test their agents with the deferred setting before moving production traffic, measure end-to-end completion time, and confirm that retries and cancellations show up on the bill as described. Developers who want the broader context on controlling agent costs can start with the Google Cloud blog index at https://cloud.google.com/blog/topics/inside-google-cloud/whats-new-google-cloud.

Sources

This article was written by Blozhik from the reporting below. Read the originals for full detail.

GoWhat’s new with Google CloudGoogle Cloud Blog
Was this useful?

Keep reading

Discussion · 0

Sign in to join the discussion. Liking does not need an account.Sign inRegister
Loading comments…
blozhikDeveloper news from across the web, explained. Every story links to its original sources.