Google's New AI Picks 100 Decisions a Second for $0.10 a Million Tokens

Google's newest model doesn't want to write you an essay. It wants to pick option B. The Decisions API returns fixed choices instead of prose, fires off up to 100 calls in under a second, and costs $0.10 per million input tokens. That's not a chatbot. That's a switchboard with a rate card.

The economics are the whole story. At a dime per million tokens, routing a decision through a frontier model costs less than the electricity on the server. So every "should we approve this?" moment in a pipeline can now go to a model instead of a hand-written rules engine. Naturally, crypto builders will aim it at the same three things they always do: allow or block a transaction, flag or clear a wallet, liquidate or hold.

Here's the hole in the pitch. The API returns choices, not confidence. The moment a decision carries real money, somebody still has to build the check that catches a wrong pick. Consequential actions still need a human, or a second model, looking over the first model's shoulder. That's the part that never makes it into the launch post.

Which seems like an even stronger argument for putting those checks on-chain and auditable. If a bot kills a $40,000 position because a classifier said so, "the model chose option C" isn't a risk disclosure. A signed record of what the model saw and what it picked is. Plenty of teams will skip that step to ship faster, and you can guess who eats the loss.

Who loses? The middleware crowd selling decision layers as a service at a healthy markup. When the raw call costs pennies, the only defensible product left is the audit trail and the exception handling. Spare me the orchestration platform pitch.

Speed was never the scarce resource. Being right at scale is, and nobody's priced that at a dime yet.