Vertex AI

What Vertex AI actually meters

Tokens, counted per model and per modality, with three qualifiers that change the bill more than the headline rate does.

First, price steps up at 200K input tokens. Gemini 3.1 Pro charges $2.00 per million input below that line and $4.00 above it, so a long-context application pays a different unit price than a short-prompt one on the same model.

Second, Google charges only for requests returning a 200 response code. Failed requests cost nothing on input or output, which is rarer than it sounds.

Third, reserved capacity meters something else. A GSU buys a fixed throughput allowance, and each model publishes burndown rates converting its modalities into it. On gemini-2.0-flash an input text token burns down as 1, an input audio token as 7, an output text token as 4. Google's worked example turns 10 queries per second into 57,000 adjusted tokens per second, divides by 3,360 per GSU, and lands on 17 GSUs.

Grounding meters separately. Gemini 3 includes 5,000 Grounding Queries a month free then bills $14 per 1,000. Gemini 2.x uses a different unit, the Grounding Prompt, with daily allowances and a $35 per 1,000 rate, or $45 for Web Grounding for Enterprise.

How credits and caching discounts work

Vertex AI has no credit wallet. What Google calls credits are billing credits applied to an invoice, and the current one is a 50% Provisioned Throughput promotion on Gemini 3.8, 3.7 and 3.6 Flash running 13 August to 31 December 2026. Those credits issue on the 7th of the following month and expire 30 days later. There's no rollover, pooling or top-up purchase.

The real discount mechanic is caching. Implicit caching is on by default on every project and takes 90% off cached input tokens with no storage charge. Explicit caching lets you control what gets cached and for how long, at 90% off on Gemini 2.5 and later and 75% on Gemini 2.0, but adds a storage charge per cached token per hour, $0.0000045 on Gemini 3.1 Pro and $0.000001 on Flash. Default TTL is 60 minutes and minimum cache sizes run 2,048 to 6,144 tokens. Cache and Batch discounts don't stack: the 90% cache hit wins.

What happens when you hit the limit

Nothing blocks by default, which is the most consequential line in the docs. A request exceeding your remaining Provisioned Throughput quota gets processed in full as an on-demand request at the pay-as-you-go rate, appearing as spillover on the monitoring dashboard.

You change that with a header. X-Vertex-AI-LLM-Request-Type: dedicated restricts you to reserved capacity and returns a 429 past it. shared bypasses the reservation entirely. The choice between a hard cap and an uncapped bill is a per-request header, not an account setting, and the default is the uncapped one.