Manage Google AI costs and caching

Google bills your company’s Cloud project for the AI used while building. Applane adds no markup and does not receive the prompts or model responses. Use this guide with your Platform and billing teams to agree on a pilot budget and allowed models.

What is billed

Item	What it covers
Input tokens	Instructions, tool descriptions and conversation context sent to the model.
Output tokens	The model’s response, including generated code.
Cached input	Repeated context read from a Vertex AI cache, at the model’s cached-input rate.
Cache storage	Keeping that context in Vertex AI for the cache’s lifetime.

Rates depend on the model and location. Check Google’s current Vertex AI pricing when estimating costs. A single test build is not a reliable monthly forecast: app complexity and follow-up changes affect usage.

Context caching

Building an app can take several rounds of AI requests. Applane caches repeated context for a build turn so it does not pay the full input rate for that context each time. The cache contains the system instructions and tool definitions, which may include app source and data-source structure used in the request.

The current builder uses a Vertex AI cache in the global location for its configured models, including models whose requests use the eu multi-region route. The cache has a one-hour lifetime and the builder tries to delete it when the turn ends. It remains a resource in your Google project; a project boundary is not a regional processing guarantee.

Turning it off

If a global cache does not fit your company’s requirements, turn Context caching off in Settings > Features. Requests are then billed without explicit caching. Review the selected model’s processing location separately.

The builder’s Costs panel reports when caching is off or unavailable. Building can continue without it.

What it needs

Caching needs aiplatform.cachedContents.create, get, list and delete in the project, alongside the model-call permissions. Review the IAM guide when using a custom role.

Review spend during the pilot

1. Limit building access to the agreed group and choose approved models in Settings.
2. Set a Google Cloud budget alert. Alerts notify the billing team; they do not cap spending.
3. Run a few representative tasks and inspect actual project usage in Cloud Billing.
4. Adjust the budget and model choices before adding more builders.