Gemini Decisions API: Fast, calibrated multimodal decisions

blog.google3 min read

Google has introduced the Gemini Decisions API powered by the new Gemini Decision 1 model, designed for fast, closed-set decisions such as tool selection, policy checks, and model routing. The model processes multimodal inputs and up to 1,000 independent questions in a single forward pass, returning probabilities and confidence scores with an end-to-end latency of 150 to 250 milliseconds. Priced at $0.10 per 1M input tokens with zero output tokens billed, the API is available through the Gemini API in Google AI Studio. Open article →

Notable comments

Discussion is broadly positive on the practicality and speed/latency/cost of Gemini Decisions for closed-set, batched classification/routing, with key technical concerns centered on calibration brittleness under task/OOD shifts and on what “decision models” are distinct from (e.g., BERT/NLI-style discriminators).

HN · richardw Shared first-hand use classifying ~10k repos and using “does this contain an entity” to pre-filter before extraction, highlighting the appeal of low-variance/mostly-deterministic behavior versus token-by-token randomness.

HN · reexpressionist Gives rigorous calibration critique: cross-entropy/logit calibration tends to work marginally under i.i.d., but breaks under covariate/task shift and OOD; ECE is not the right metric for conditional branching reliability and would require specialized recalibration with representation constraints.

HN · skapadia Provides concrete application patterns (task classification, guardrail/policy checks, and using a rubric-like decision query as a fast cross-encoder/reranker) explaining how to structure inputs for closed-set decision workflows.

More on this topic