The Gemini (Vertex) API now provides restricted entry to grounding queries and returns the search phrases within the [‘groundingMetadata’] object. That’s thrilling information, in concept, nevertheless it seems to be messy in apply. We encountered two main points:
1. Gemini is grasping about grounding
Even operating Gemini 3.5 Flash with a [MEDIUM] or greater “considering” degree, it’s uncommon to get greater than 1-2 grounding queries. In lots of instances, you’ll see none in any respect. In the event you’ve used AI Mode a lot, you realize it nearly all the time runs a number of grounding queries. We had been capable of nudge the system directions and push Google to generate 4-6 grounding queries. Whereas these queries are “actual” (Google truly ran the searches), it’s honest to say that we cheated a little bit bit. AI Mode and precise question fan-out would possibly behave very in another way.
2. Google is grasping in your cash
On high of token prices, grounding queries presently price (as of August 2026) $14 per 1,000 queries. This isn’t a per-prompt price. In our case, every immediate ran about 5 grounding queries, so grounding knowledge price one thing like $7 per 100 prompts. That is orders of magnitude greater than the token price, and at enterprise-scale, it provides up quick. In the event you’re utilizing Third-party instruments that run by yourself API keys, hold a detailed eye on prices.
