SXF RESEARCH EXPLAINER
What the technical evidence means.
TECHNICAL ANALYSIS / 01
What is Gemini Decision 1, and when should you use it?
Gemini Decision 1 is designed for closed choices: selecting a tool, deciding whether a document matches policy, grading a response against a rubric or routing a task to a more capable model. It does not replace a generative assistant that needs to explain findings, write code or produce an answer grounded in multiple passages.
A useful distinction is that a decision call can say which known option is most likely while your application supplies the meaning of those options and controls the next step. An incorrect routing choice can still trigger a risky action, so permissions must be enforced by the application rather than left to the model's label.
Check primary-source evidence ↗
TECHNICAL ANALYSIS / 02
How much does Gemini Decisions API cost per request?
Google publishes a list price of $0.10 per million input tokens and no billed output tokens. The input can be reused across many questions in the same request, but the instructions, labels and criteria describing those questions also count toward input usage. The claim that 1,000 questions fit in one request does not mean every question is free.
SXF's illustrative budget: a request with 10,000 shared-input tokens and 5,000 question-and-criteria tokens totals 15,000 billed input tokens. At the posted list rate, the token component would be $0.0015. That is a transparent calculation, not a billed request or a guarantee about tokenization of video, audio and PDF inputs.
Check primary-source evidence ↗
TECHNICAL ANALYSIS / 03
Latency, calibration and what the benchmarks actually measure
The announcement reports 50–100 ms median compute and 150–250 ms end-to-end latency across a tested slice of multimodal inputs, contexts up to 16K and question counts around 1 to 100. That must not be generalized to a maximum-size 256K input with 1,000 questions. Network distance, modality preprocessing and batch composition affect real latency.
Google describes lower calibration error on its evaluation suite. Before use, measure accuracy, false positives and confidence reliability on representative cases. For example, collect human-labelled examples in each supported language or document type, bucket predicted confidence and compare predicted scores with observed error rates. A confident result can still be wrong.
Check primary-source evidence ↗
TECHNICAL ANALYSIS / 04
SXF implementation pattern for a safer classification agent
Define an allowed list of decisions, explicit exclusion cases and a fallback. Record the decision prompt version, anonymized category, selected option, score, eventual human outcome and time spent. Reserve automatic execution for well-tested low-risk actions, escalate ambiguous cases and fail closed on unknown input types.
For a policy checker, start with triage rather than enforcement: classify whether a case deserves human review, but do not use that score alone to deny an account or block a payment. Compare total decision cost and error-adjusted handling time against a conventional rule system and a generative model on identical tasks.
Check primary-source evidence ↗