Single pass decisions cut the cost of agent routing

Most agent architectures spend a surprising amount of compute asking large language models to make simple categorical choices. Forcing a generative decoder to emit text one token at a time just to choose between three downstream tools or output a binary classification adds latency and invites formatting failures. Google's Gemini Decisions API, powered by their Gemini Decision 1 model, strips out text generation entirely, scoring up to 1,000 closed-set questions across text, audio, video, and documents in a single pass.

The operational details matter here. Pichai reports that the API supports yes-or-no decisions, multiple choice with up to 255 options, and rubric scoring, priced at ten cents per million input tokens with zero charge for output tokens. Because the model does not generate freeform tokens, it sidesteps the standard failure modes of constrained decoding, such as malformed JSON or prompt drift where a model hallucinates an unlisted option. Moving scoring into the forward pass also cuts down the round trips that drag down agent loops.

If Google's reported calibration gains hold up in production, this offers a clean blueprint for how agent backbones ought to divide labor. Frontier generative models are valuable for synthesis and complex drafting, but using them as glorified classifiers has always been expensive overkill. Offloading routing and deterministic grading to dedicated decision models makes agentic architectures cheaper and considerably easier to verify.