LLMs

Google Launches Gemini Decision 1 for Fast, Calibrated Classification Across Multimodal Inputs

Google DeepMind releases Gemini Decision 1 via the Decisions API, enabling sub-250ms scoring of up to 1,000 yes/no and multiple-choice questions on text, images, audio, video, and PDFs for $0.10/MTok.

Last verified:

Gemini Decision 1: Purpose-Built for Classification at Scale

Google DeepMind announced the Gemini Decisions API on November 11, powered by Gemini Decision 1, a specialized model engineered for fast, bounded-choice classification tasks. According to the Google AI Blog, the API evaluates up to 1,000 independent yes/no, multiple-choice, and rubric-scoring questions in a single forward pass, accepting multimodal input across text, images, audio, video, and PDFs (up to 256K tokens total) with latency of 150–250 milliseconds end-to-end and pricing of $0.10 per million input tokens.

The model’s architecture diverges sharply from generative inference: instead of decoding output tokens one-at-one, Gemini Decision 1 scores all candidate options in a single forward pass, eliminating the sequential cost of autoregressive generation. This design choice directly addresses the latency and cost profile of production decision systems—tool routing, policy checks, model routing, context grading—where the output space is known and bounded.

Performance and Calibration Benchmarks

According to the Google AI Blog, Gemini Decision 1 leads across a suite of agent-routing, reasoning, ranking, and multimodal understanding benchmarks with the lowest calibration error. The model maintains median inference compute of 50–100 milliseconds and end-to-end latency of 150–250 milliseconds across multimodal inputs and contexts up to 16K tokens, remaining nearly flat as batch size grows from 1 to 100 questions per request. Critically, since classification scores are not billed as output tokens, the model costs approximately 5 times less than Gemini 3.5 Flash-Lite on shared benchmarks for comparable tasks.

Each answer includes a confidence score (0–1), allowing developer code to route ambiguous classifications to fallback systems or human review without additional inference overhead.

Multimodal Evaluation in One Pass

The API integrates three decision types in a single request: yes/no checks with optional true/false criteria; multiple-choice selection from 2–255 candidates; and rubric-based scoring. The model processes all input modalities—text, images, audio, video, PDF—natively in its forward pass, permitting developers to combine modalities within a shared input (for example, pairing a call recording with account metadata or a scanned PDF claim with damage photographs). Developers can tune image and video resolution independently to balance visual detail against token consumption.

Pricing and Availability

At $0.10/MTok with zero output token billing, Gemini Decision 1 becomes economical for high-volume classification workloads. The Decisions API is available through the Gemini API in Google AI Studio, immediately accessible to developers via the gemini-decision-1 endpoint.

Why This Matters

This release targets a concrete economic and latency pain point: production systems that currently invoke a general-purpose LLM to select from a known set of options (tools, models, policies) now have a 5x cheaper, sub-250ms alternative. Teams deploying agent-based systems, content moderation, model routing, or claim triage benefit most immediately. The emphasis on calibration—reported as lowest-in-class—signals Google’s recognition that confidence scores, not raw accuracy, determine whether downstream fallback routing remains practical at scale. If the benchmark claims hold under independent reproduction, Gemini Decision 1 will likely shift cost-sensitive production classification workloads away from general-purpose models, intensifying pressure on inference pricing across the industry.

Frequently Asked Questions

What is the Gemini Decisions API and why does it exist?

The Decisions API is a purpose-built endpoint for fast, closed-set classification tasks—like tool routing, policy checks, and context ranking—that scores multiple candidate options in a single forward pass rather than generating text token by token. It is optimized for latency and cost when the output space is bounded.

How does Gemini Decision 1 pricing differ from other Gemini models?

Gemini Decision 1 charges $0.10 per million input tokens with zero output token billing, making it approximately 5x cheaper than Gemini 3.5 Flash-Lite on comparable tasks, since classification scores do not consume output tokens.

What modalities does Gemini Decision 1 support?

The model natively processes text, images, audio, video, and PDFs in a single forward pass, with up to 256K total token context and support for tunable image and video resolution to balance visual detail against token cost.

What decision types can the API evaluate in one request?

The Decisions API supports three decision types in a single request: yes/no questions with optional criteria, multiple-choice selections from 2–255 options, and rubric-based scoring. Every answer includes a confidence score (0–1) for fallback routing.

#gemini #api #multimodal #inference #pricing #google-deepmind