How to cut your Vertex AI API bill

Mode

Both: automatic caching by default, plus explicit breakpoints you control

Implicit caching (in-memory plus durable tier) on by default for all projects; explicit caching via cached-content resources on Gemini 3.x and 2.5 models.

Cache lifetime

Implicit: a cached prefix expires no later than 24 hours after creation, most entries earlier. Explicit: default 60 minutes, minimum 1 minute, no maximum.

Minimum prompt

Gemini 3 (with exceptions) and Gemini 4 models: 4,096 tokens. Gemini 3.0 Flash Preview, 3.1 Pro Preview, 3.7 Flash, and 3.8 Flash (implicit caching only): 6,144 tokens. Gemini 2 family: 2,048 tokens.

Cache write cost

Cache creation billed at the standard input token price (no premium)

Cache read cost

Implicit: 90% discount on cached tokens. Explicit: 90% discount on Gemini 2.5 or later, 75% on Gemini 2.0.

Storage cost

Explicit caching only: storage costs based on how long caches are stored. No storage cost for implicit caching.

How to enable

Implicit: on by default (can disable durable tier or all implicit caching). Explicit: create a context cache via the Gemini Enterprise API and reference it by resource name.

From the docs

“Implicit caching: Automatic caching enabled by default that provides cost savings when cache hits occur. Implicit caching includes in-memory caching and durable caching.”

“Explicit caching: Manual caching enabled using the Gemini Enterprise API, where you explicitly declare the content you want to cache and whether or not your prompts should refer to the cached content.”

“A cached prefix expires no later than 24 hours after it's created, and most entries are evicted earlier.”

“Update a context cache's expiration time (Time to Live, or TTL) past the default 60 minutes.”

“| Minimum time before a cache expires after it's created | 1 minute |”

“Gemini 3 (with exceptions) and Gemini 4 models: 4,096 tokens”

“Gemini 3.0 Flash Preview, 3.1 Pro Preview, 3.7 Flash, and 3.8 Flash (implicit caching only): 6,144 tokens”

“For both implicit and explicit caching, you're billed for the input tokens used to create the cache at the standard input token price. For explicit caching, there are also storage costs based on how long caches are stored. There are no storage costs for implicit caching.”

“Implicit caching provides a 90% discount on cached tokens compared to standard input tokens.”

“On Gemini 2.5 or later models, this discount is 90%; on Gemini 2.0 models, this discount is 75%.”

Discount

50% off

50% discount vs real-time; does not stack with the 90% implicit cache discount (cache discount takes precedence)

Window

24-hour turnaround target; queue up to 72 hours before expiry; incomplete jobs cancelled 24 hours after they start running (charged only for completed requests)

Limits

Up to 200,000 requests per job; 1GB file limit with Cloud Storage input; no predefined quota (shared pool); no Provisioned Throughput or explicit caching

Endpoint

Batch inference job (Cloud Storage or BigQuery input)

Stacks with caching

No: you get one discount or the other, not both

“However, the discounts for cache and batch don't stack. The 90% cache hit discount takes precedence over the batch discount.”

From the docs

“Batch processing is offered at a 50% discounted rate compared to real-time inference”

“However, the discounts for cache and batch don't stack. The 90% cache hit discount takes precedence over the batch discount.”

“The job remains in queue for up to 72 hours before it expires.”

“Most jobs complete within 24 hours after it starts running (not counting the queue time). After 24 hours, incomplete jobs are cancelled, and you'll only be charged for completed requests.”

“A single batch job may include up to 200,000 requests. If you are using Cloud Storage as input, there is also a file size limit of 1GB.”

“Batch inference doesn't support Provisioned Throughput, explicit caching or RAG.”

• Provisioned Throughput (dedicated, assured capacity)

  “For workloads that need dedicated, assured capacity, see Provisioned Throughput.”

This site does not track per-model Vertex AI prices. Use the sources below for the current rate card.