What does Applane cost on Vertex AI?

Every model call Applane makes runs on Vertex AI in your own GCP project, with the builder’s own account. Google bills your project at its list price for the model in use. Applane never meters, marks up or sees those calls: no prompt and no model output ever reaches Applane.

What is billed

Item	What it is	How it is priced
Input tokens	The system prompt, the tool declarations and the conversation so far, sent on every round of a build turn.	Per million tokens, at the model’s input rate.
Output tokens	What the model writes back: code, tool calls, explanations.	Per million tokens, at the model’s output rate, which is higher than input.
Cached input tokens	The part of the input served from a context cache instead of being re-sent.	About 10% of the input rate.
Cache storage	Holding a context cache in Vertex AI.	Per million tokens per hour, for as long as the cache exists.

Apps Script, Sheets, Drive and Docs calls have no per-call charge for Workspace users.

The measured number

In our own testing, a full day of building used $0.59 of Gemini tokens, with no caching. Your usage depends on how many builders you have and how much they build. The budget alert is the control.

Context caching

A build turn is a loop: the model reads, calls a tool, reads the result, and repeats until the app is done. The system prompt and the tool declarations are identical on every round, and they are most of the input. Re-sending them each round is what makes a long turn expensive.

The extension avoids that with Vertex AI explicit context caching (cachedContents). At the start of a turn it stores the system prompt and tool declarations in a cache, then references that cache on every following round. Each round pays the cached rate for that part instead of the full input rate. Measured on our build benchmark, this cuts the model cost of a build turn by 79 to 90%; the exact saving depends on how many rounds the turn takes.

Where the cache lives	In your project, in Vertex AI’s global location. Vertex serves context caches only from global, even for requests routed to eu, so the cached system prompt (the app’s source, the sheet structure, the tool declarations) sits in a global Vertex resource for up to one hour. It never leaves your project.
What it holds	The system prompt and tool declarations for that turn. No sheet data, no app source beyond what the turn already sends.
How long it lives	The cache has a one-hour time to live. The extension deletes it when the turn ends, so storage is billed for minutes, not hours.
Who can read it	Only the builder who created it, through the same IAM the model call uses.

Caching changes nothing about what the model sees or answers. It only changes how the input is billed.

Turning it off

If a global cache does not fit your data-residency rules, the Applane admin can turn Context caching off under Settings > Features. Every request then stays in the model’s region and is billed at the full input rate, and the Costs panel shows “Context cache: off — disabled for this company”. An EU-resident alternative exists at a price: gemini-3.5-flash in europe-west3 caches normally, at about twice the input rate of the default model.

What it needs

The four aiplatform.cachedContents.create, get, list and delete permissions, which roles/aiplatform.user already includes. Companies that grant a custom role instead must add them, plus aiplatform.endpoints.predict. Details on the IAM page.

Without them the extension still works: every turn is billed uncached, and the Costs panel in the extension shows “Context cache: off”.

Keeping the bill predictable

• Grant roles/aiplatform.user to the builders group, not to everyone, so only builders can generate Vertex spend. IAM
• Set a budget alert on the project.
• Limit the models builders may pick under Settings in the Applane console.
• Read the per-user cost in Cloud Billing reports; every call is attributed to the builder’s own account.