Google has introduced Gemini Decision 1, a developer API built to make fast, structured judgments for AI agents and other applications, rather than generate long-form responses.

The API can evaluate text, images, audio, video and PDFs, and assess up to 1,000 questions against shared material in a context window of up to 262,144 tokens; supported formats include yes-or-no checks, multiple choice with up to 255 options, and criteria-based scoring.

That batch approach is designed for bounded decisions such as routing requests, checking policies, selecting tools or model tiers, grading retrieved information, and flagging uncertain cases for human review, potentially avoiding repeated calls to a general-purpose model.

Gemini Decision 1 returns concise, structured answers with confidence scores, while larger models or application rules can handle complex reasoning and actions.

Its multimodal inputs let developers combine materials such as call recordings and customer data or damage reports and photos, potentially avoiding separate transcription and scoring steps.

Questions in a batch are evaluated independently, so developers must build application logic or send follow-up requests when one decision depends on another; conversation history must also be supplied as shared text when relevant.

The API has important guardrails and limitations: multiple-choice questions always select one of the listed options, so Google recommends including an “other” choice or checking scope separately, and safety-filtered requests return errors rather than low-confidence judgments.

Google’s reported benchmark results are company measurements, not guarantees of real-world performance, and independent testing has not established how well the confidence scores will be calibrated on customers’ data.

Before production use, developers should test labeled examples from real traffic—including rare cases and poor-quality media—set confidence thresholds and escalation paths, and measure errors, review rates and end-to-end latency.

Developers remain responsible for defining sound criteria and keeping human or secondary review for high-impact decisions; numerical comparisons, permissions and other consequential checks should stay under application controls.

Possible applications include risk assessment and credit scoring, but Omdia’s Lian Jye Su said organizations should assess performance in real operating conditions and fit with existing environments, not rely only on published benchmarks.

Google offers a no-cost tier for prototyping, demo apps and Python, JavaScript and REST guides; integrations are rolling out across developer platforms, with Google Cloud support planned for later.