GCP

Google Cloud: 50% Provisioned Throughput Credit for Gemini Flash Through 2026; GKE 1.37 in Rapid

Google Cloud gives a 50% Provisioned Throughput credit for Gemini Flash through Dec 31, 2026, and adds GKE 1.37 to the Rapid channel, changing TCO now.

September 10, 2026·3 min read·AI researched · AI written · AI reviewed

Google Cloud just handed platform teams a narrow, time‑boxed economic lever that will change model placement decisions: effective Aug 13, 2026 through Dec 31, 2026, Gemini 3.8/3.7/3.6 Flash Provisioned Throughput spending is eligible for a monthly billing credit equal to 50% of net eligible spend. In plain terms: if you plan to run high‑volume, latency‑sensitive inference on Gemini Flash with Provisioned Throughput, your effective token cost this quarter can be cut in half — but only while the credit exists.

The numbers matter. Google’s pricing catalog lists introductory per‑million‑token rates for Flash variants and shows higher list prices after Jan 1, 2027; consult the official pricing page for the exact per‑token amounts and tier mappings. The documentation also separates cached versus non‑cached input tokens and calls out context caching and storage charges; for systems with heavy context reuse those cache read and tokens‑hour storage costs can be material. Trackers such as BenchLM and similar sites now publish synchronized model IDs, context limits, and the per‑token rates Google documents, which makes cost comparisons straightforward if you validate against Google’s catalog.

This is not a small detail for architecture: treat the 50% Provisioned Throughput credit as a customer acquisition subsidy, not a new steady‑state price. It makes Provisioned Throughput far more attractive for bursty, low‑latency services during the promo window — you’ll see lower per‑request variance and predictable SLOs for less cash outlay. But Jan 1, 2027 matters. If you bake provisioning or architectural choices around promotional economics, you’ll wake up to a list‑price cliff unless you actively rearchitect, renegotiate, or accept higher run rates.

Another gotcha: cached tokens. Google explicitly differentiates cached reads from fresh model input. For conversational systems with large context reuse — evaluation, retrieval augmented generation, or multimodal pipelines with repeating prompts — cache behavior and cache read/storage charges can dominate cost. Architects must instrument not only raw token counts but cache hit rates and tokens‑hour metrics. Long context retention trades saved latency and reduced output token usage for a standing storage bill.

Operationally, this is good for Google — incentivize large customers to try Provisioned Throughput, then let a smaller steady price retain them. It’s also exactly the kind of move that will suck teams into under‑modeling the post‑promo TCO. My take: treat the credit as a limited‑time migration window. Prototype and prove value under the promo, then run a 2027 budget scenario immediately.

While pricing grabbed the headlines, the platform updates include an operationally relevant GKE note: Kubernetes 1.37 images are available in the Rapid channel, and Google’s release notes indicate some 1.36.x and 1.37.x variants are being used as Rapid defaults in managed workflows. That affects upgrade cadence, node pool creation defaults, and multi‑cluster topology plans — especially for teams harmonizing GKE Rapid with strict regulatory baselines. If you use Rapid channel clusters for CI/perf testing or edge fleets, expect 1.37 images to be the path of least resistance for new cluster creation and adjust your upgrade testing matrix accordingly.

Two last points engineers should bookmark: cost trackers like BenchLM now list the Flash model IDs and context lengths that Google documents, which makes automated cost comparisons (online vs batch, Flash vs other Gemini releases) easier; and Google’s broader posts this week emphasize container security and distributed systems patterns that still inform how you should host inference — the economic incentives only amplify existing operational tradeoffs.

If you’re responsible for AI platform TCO, do this now: run a side‑by‑side estimate of Provisioned Throughput with and without the 50% credit, model cache hit scenarios with tokens‑hour storage, and plan for a Jan 1, 2027 price reality. This credit is a smart short‑term play by Google, but it’s a trap for the lazy: promotional pricing is a launchpad, not an operating contract. For background on Gemini Flash access and gating, see my earlier note on Gemini Flash Cyber.

Prediction: the next six months will see a cluster of migrations to Provisioned Throughput for latency‑sensitive services, followed by a reckoning in Q1 2027 when teams either optimize caches aggressively or move cold workloads to cheaper batch inference. If you don’t have a plan for that transition today, you will be paying twice as much for the same traffic next year.

Sources

google-cloudgemini-apigke-1-37ai-pricing
← All articles
GCP

Cloud Run Instances (Preview): long-lived, addressable workloads and per-instance pricing

Cloud Run Instances (Preview) expose long‑lived, addressable workloads with per‑instance billing (1 vCPU + 1 GiB steady cost). Rethink always‑on agents.

Sep 9, 2026·3mcloud-rungke
GCP

Cloud Run adds NVIDIA L4 GPU support; Cloud Functions-to-Cloud Run upgrade tool GA

Cloud Run adds NVIDIA L4 GPU support with managed drivers; a Cloud Functions-to-Cloud Run upgrade tool is GA—simplifying serverless GPU inference migrations.

Sep 8, 2026·3mgcpcloud-run
GCP

GKE 1.36: Dataplane V2 Emits CNI cniVersion 1.1.0 — Upgrade Risk for CNI Plugins

Dataplane V2 in GKE 1.36 emits CNI configs with cniVersion 1.1.0. Plugins lacking 1.1.0 semantics can fail to set up pod networking during upgrades — validate.

Sep 6, 2026·3mgkekubernetes