Google DeepMind Launches Gemini Decisions API With Specialized 'Decision 1' Model for Calibrated Multimodal Judgments

Google DeepMind on November 11 released the Gemini Decisions API, a new endpoint powered by a purpose-built model called Gemini Decision 1 (gemini-decision-1). The API is designed for closed-set decisions — yes/no, multiple choice, and rubric scoring — rather than open-ended text generation. It accepts multimodal inputs and returns calibrated probabilities and confidence scores in a single forward pass.

What the API Does

The Decisions API evaluates up to 1,000 independent questions against one shared input of up to 256,000 tokens. Inputs can combine text, images, audio, video, and PDFs. Each question uses one of three types:

• noul (yes/no): returns a probability between 0.01 and 0.99 that a condition holds, with optional criteria for what "yes" and "no" mean.
• choice: selects one option from 2 to 255 candidates and returns normalized probabilities for every option.
• score: rates the input on an ordered rubric of 2 to 255 levels and returns a probability-weighted score.

Every answer includes a confidence score from 0 to 1, intended to let developers route low-confidence cases to a fallback or human reviewer. Questions are scored in parallel and cannot see each other. The shared input is billed once per request; each question adds only its instruction and criteria tokens. There are no output tokens, so output billing is zero. [1][6]

Performance Claims and Benchmarks

Google's announcement includes a benchmark table comparing Gemini Decision 1 against three rivals: TypeSafe Jev 1.14, GPT-6 Luna (OpenAI Decisions), and Gemini 3.5 Flash-Lite. All figures are self-reported by Google; refusals and safety blocks count as incorrect.

Gemini Decision 1 leads on seven of nine benchmarks, including JevBench (92.2%), New Yorker Caption (79.7%), PubMedQA (79.8%), Global-MMLU-Lite (86.0%), MMAU audio (71.8%), and Video-MME (69.1%). It trails Jev 1.14 on Mind2Web Multimodal (86.2% vs 86.8%) and When2Call (79.9% vs 82.0%). Its lowest score is 58.7% Macro-F1 on Amazon ESCI.

The clearest gap is calibration: Google reports a top-1 expected calibration error (ECE) of 0.0250 for Decision 1, versus 0.0494 for Jev 1.14, 0.0721 for GPT-6 Luna, and 0.2104 for Flash-Lite. Calibration matters because developers set automation thresholds on these confidence scores. No independent verification of these numbers has been published. [1][3]

On cost, Google says Decision 1 is nearly 5x cheaper than Flash-Lite across its shared benchmark suite ($0.88 vs $4.14 per million requests at 2,000 shared input tokens). However, TypeSafe Jev 1.14 lists a lower input price ($0.042 per 1M tokens) and a lower total on the same suite ($0.68). Google does not yet support prompt caching or a Batch API, so repeated context across requests is billed in full. [1][3]

Latency and Architecture

Google reports median model compute of 50–100 ms and end-to-end latency of 150–250 ms, nearly flat from 1 to 100 questions per request for inputs up to 16K tokens. The speed comes from skipping token-by-token decoding: the model scores candidate options in one forward pass. These are vendor figures; the announcement does not publish test conditions (region, input size distribution, percentile data). Independent latency tests have not yet appeared. [1][6]

Internal Deployments

Google says several teams are already using the Decisions API in production:

• Gemini Notebook: routes queries to model tiers and reasoning levels, cutting latency for everyday queries by up to 40% while reserving deep reasoning models for complex tasks.
• Stitch (rapid canvas): generates playable wireframes in seconds and applies voice or tap edits almost instantly.
• CC (Google Labs): selects emoji reactions from more than 200 options in a fraction of a second.
• Google Search: uses it as a safety layer to detect adversarial behavior involving Model Context Protocol (MCP) tools. [1][5]

Availability and Ecosystem

The API is live on the Gemini Developer API (v1beta) and requires the google-genai SDK 3.2 or later. A no-cost tier is available for prototyping in Google AI Studio. Python, JavaScript, and REST guides are in the documentation. Integrations are rolling out across OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. Support in Gemini on Google Cloud is listed as "coming soon." [1][6]

OpenRouter shows the model hosted by Google AI Studio at $0.10/M input tokens, zero output, with 0.23s latency and 100% uptime over three days. The model identifier is google/gemini-decision-1 on OpenRouter, which uses its own Decisions API endpoint rather than the OpenAI-compatible chat endpoint. [2]

Competitive Context

OpenAI launched its Decisions API in October, powered by GPT-6 Luna. OpenAI claims decisions up to 10x faster than its Responses API, up to 400 questions per request, and pricing starting at $0.10/M input tokens (uncached). OpenAI supports prompt caching; Google does not yet. TypeSafe Jev 1.14 offers a lower input price ($0.042/M) but is text-only and has a smaller context window (64K per request, 262K with opt-in header). Gemini Decision 1 supports five input modalities natively. [1][3][5]

What Remains Unconfirmed

• Calibration quality: Google states the model returns "calibrated probabilities" and reports the lowest ECE in its table, but no independent evaluation on real tasks has been published. A stated feature is not a measured result.
• Per-modality accuracy: The announcement does not break down performance by text, image, audio, video, or PDF. Audio and video questions are often harder for any model.
• Rate limits, regional availability, SLAs, and data-retention terms specific to this model are not detailed in the announcement.
• Free-tier limits for the AI Studio no-cost tier are not specified.
• Token accounting for media: Google has not published how images, audio, or video convert into tokens for this model, so developers cannot estimate the cost of, say, a one-hour video by arithmetic alone.

Bottom Line

Gemini Decision 1 is a specialized, hosted model for high-volume, low-stakes judgment calls — routing, tagging, moderation, policy checks, and first-pass grading. By confining itself to closed-set decisions, it trades flexibility for price, speed, and structured output that software can act on directly. At $0.10 per 1M input tokens with zero output billing, the cost case is clear on paper.

The proof case is still open. Benchmark leadership and calibration claims are self-reported. Developers who test on their own labeled data in the coming weeks will learn more than any launch post can tell them.