What Is Gemini Decision 1? Specs, Cost & Speed

Scoring 1,000 Questions at $0.10 per Million Tokens: A First Look at Gemini Decision 1

Gemini Decision 1 is a multimodal decision model developed by Google DeepMind that scores up to 1,000 closed-ended questions against a shared input in a single forward pass without generating text tokens. Officially launched on November 11, 2026, through the Gemini Decisions API in Google AI Studio, the model accepts text, audio, video, images, and PDF documents across a 256,000-token context window. It returns calibrated probability distributions and confidence scores in 150 to 250 milliseconds end to end, billed at $0.10 per 1 million input tokens with zero charge for output.

Key Takeaways

• Architecture Shift: Gemini Decision 1 replaces autoregressive token-by-token text decoding with a single forward pass that evaluates candidate answers directly.
• Zero Output Fees: The model bills exclusively for input tokens at $0.10 per 1 million tokens; output token consumption is permanently zero.
• High-Density Batches: A single API call can evaluate between 1 and 1,000 independent questions over the same shared multimodal context.
• Input Flexibility: Native support covers mixed payloads including text, image files, audio tracks, video clips, and PDFs up to 256,000 tokens (documented as 262,144 tokens in API reference tables).
• Fast Turnaround: Typical end-to-end response times range between 150 and 250 milliseconds, with model computation taking 50 to 100 milliseconds on inputs under 16,000 tokens.
• Dual Performance Profile: Google tests show Gemini Decision 1 beating TypeSafe's Jev 1.14 on 5 of 7 shared suites, though independent indices place its aggregate performance near the middle of dedicated decision benchmarks.

What Is Gemini Decision 1?

Gemini Decision 1 is Google's entry into the dedicated decision model category, a model class optimized for categorical classification, routing, and numerical scoring rather than open-ended dialogue. Standard large language models solve structured classification by generating JSON or text tokens, which requires token decoding loops, consumes output budgets, and introduces parsing failures. Gemini Decision 1 bypasses text decoding altogether.

Announced by Google DeepMind, the model operates behind the gemini-decision-1 identifier under the v1beta Decisions API. Instead of interacting through standard chat endpoints, developers query the model using dedicated SDK methods like client.decisions.create available in google-genai version 3.2 and later.

Source: @shao__meng

The release arrived alongside a broader expansion of Google's frontier portfolio, including our analysis of Gemini 4 Carbon and Claude Opus 5.5. While models like Carbon target deep multi-step software engineering, Gemini Decision 1 is designed specifically as an ultra-fast control-plane model. It evaluates evidence, calculates bounded probabilities, and supplies downstream programmatic workflows with deterministic scores.

Gemini Decision 1 at a Glance

Specification	Detail
Developer	Google DeepMind
Model Identifier	gemini-decision-1
Release Date	November 11, 2026 (Preview / v1beta)
Model Type	Multimodal Decision / Scoring Model
Input Modalities	Text, Images, Audio, Video, PDFs
Context Window	256,000 tokens (262,144 tokens API limit)
Max Questions per Call	1,000 independent questions
Output Token Count	0 tokens (scores returned via metadata)
End-to-End Latency	150 to 250 ms (50 to 100 ms compute on ≤16K context)
Input Pricing	$0.10 per 1M tokens
Output Pricing	$0.00 (Free)
Free Tier Quotas	200 RPM / 50,000 TPM / 10,000 requests per day
API Endpoints	Google AI Studio (POST /v1beta/decisions), Gemini API
Enterprise Availability	Google Cloud Vertex AI (listed as coming soon)

How Gemini Decision 1 Works: Single-Pass Probability Scoring

Traditional autoregressive classification forces a language model to sample tokens sequentially from a vocabulary distribution until a valid structured string is complete. Gemini Decision 1 alters this execution flow by treating classification as a direct candidate-scoring problem.

The Single Forward Pass

When a payload enters the Decisions API, the model ingests the shared multimodal context alongside up to 1,000 questions. The shared input is processed through the attention layers once. The model then scores the candidate outputs for all submitted questions in parallel. Because the questions cannot attend to each other's candidate answers, every question remains isolated: adding, reordering, or removing a question does not change the scores of the remaining items.

Because no auto-regressive generation occurs, the system eliminates decoding latency. According to launch metrics from Google AI Studio, model computation remains flat between 50 and 100 milliseconds whether answering 1 question or 100 questions, provided the shared input context remains under 16,000 tokens.

Three Native Question Types

Developers structure their queries using three primitive formats rather than free-form text prompts:

1. noul (Binary Yes/No): Checks whether a condition is satisfied. The model returns a normalized probability ranging strictly between 0.01 and 0.99, preventing synthetic overconfidence spikes. Optional true_description and false_description fields clarify ambiguous boundary conditions.
2. choice (Multiple Choice): Selects a single winner from a predefined list of 2 to 255 options. The API requires a map of option identifiers to semantic descriptions. The model outputs the winning key along with normalized probabilities across all submitted choices that sum to 1.0.
3. score (Ordered Rubric): Evaluates the context against an ordinal progression containing between 2 and 255 levels (2 to 10 levels recommended for optimal calibration). The return payload supplies a continuous weighted score alongside individual probabilities for each level.

Calibration and Confidence Scores

Every evaluated question returns both a value (such as the selected option or probability) and a separate confidence score bounded between 0.0 and 1.0.

"When a workflow or agent just needs to make a quick call, the Decisions API returns calibrated confidence scores at $0.10 per 1M input tokens, making it nearly 5x more cost-efficient across the shared benchmark suite than Gemini 3.5 Flash-Lite."

This confidence metric functions as an automated circuit breaker. Programmers can write conditional logic that accepts automated routing if the confidence score clears 0.85, but triggers human escalation or transfers the context to a heavier reasoning model if confidence drops below a set threshold.

What You Can Do With Gemini Decision 1

Production systems rarely require natural language paragraphs when directing system traffic. Gemini Decision 1 fits into latency-critical, high-volume control planes where deterministic numbers are needed.

Dynamic Agent and Model Routing

In large-scale agent architectures, routing every query through a frontier model creates unacceptable cost and latency overhead. Google reported integrating Gemini Decision 1 into Gemini Notebook to inspect incoming queries and determine whether they require basic retrieval, code execution, or advanced reasoning tiers. Google claims this dynamic dispatch reduced everyday query latency by up to 40%. Engineers exploring small models for high-throughput routing will recognize this trade-off from our breakdown of Meta's Muse Ember agent model.

Multimodal Triage and Document Verification

Because Gemini Decision 1 accepts images, video, audio, and PDFs natively, teams can validate mixed media in a single network round-trip. For example, an insurance application can submit a claim PDF, damage photographs, and an adjuster voice note simultaneously. Using a single request, the system can execute distinct evaluations:

• A noul check to verify whether damage photos match the text description.
• A choice operation selecting the claim severity category across 10 possible policy paths.
• A score metric assessing fraud risk on an ordered scale from 0 to 5.

Because the shared media context is tokenized and charged only once, evaluating dozens of compliance criteria adds zero additional input media fees.

Real-Time Safety Layers and Tool Call Filtering

Deployments running Model Context Protocol (MCP) servers or autonomous web browsing agents can place Gemini Decision 1 in front of external tool execution. Google Search deployed the model internally to inspect proposed tool parameters and detect adversarial injection vectors before database execution. The 150 to 250 millisecond response window allows safety gates to function without stalling end-user streams.

How Gemini Decision 1 Compares

The release of Gemini Decision 1 follows an industry push toward dedicated evaluation endpoints, competing directly with TypeSafe's Jev, OpenAI's Decisions API (evaluated with GPT-6 Luna), and dedicated models like Drex Max.

Source: @kimmonismus

Google Internal Benchmarks vs. Independent Indices

Public evaluations show two distinct pictures depending on whether you consult Google's internal suite or third-party composite benchmarks.

Evaluation Metric / Suite	Gemini Decision 1	TypeSafe Jev 1.14	OpenAI Decisions (GPT-6 Luna)	Drex Max v1.1
JevBench Top-1 Accuracy	92.2%	90.0%	90.9%	Not reported
Top-1 ECE (Calibration)	0.0250	Not reported	0.0721	Not reported
When2Call Tool Routing	79.9%	82.0%	Not reported	Not reported
Amazon ESCI Ranking	58.7%	Not reported	54.7%	Not reported
Artificial Analysis Decision Index	66 (#10 of 19)	66.23	60.0	72.62 (#1)
Time per Task (AA Index)	211 ms	Not reported	Not reported	Fast (unspecified)
Cost per 1M Tasks (AA Index)	$345	Not reported	Not reported	$103
Direct Input Token Pricing	$0.10 / 1M	$0.042 / 1M	Variable	Custom
Context Window Size	256,000 tokens	Restricted	Restricted	Restricted

In Google's published evaluations, Gemini Decision 1 bested TypeSafe's Jev 1.14 on 5 of 7 shared benchmarks, showing notable strength in Expected Calibration Error (Top-1 ECE of 0.0250 versus 0.0721 for GPT-6 Luna). Lower calibration error means the model's reported probabilities align closely with true mathematical accuracy. However, in tool calling via When2Call, Gemini Decision 1 scored 79.9%, trailing Jev 1.14's 82.0%.

Independent evaluations published on Artificial Analysis tell a more measured story. On their Decision Index v2.1, Gemini Decision 1 scored 66 points out of 100, ranking #10 across 19 tested models and matching the category median. The benchmark measured its run cost at $345 per 1 million compound evaluation tasks, ranking it 16th for cost efficiency behind specialized options like Drex Max v1.1 ($103 per 1M tasks).

"While Google claims a 5x cost reduction against generative Flash-Lite endpoints, independent indices place Gemini Decision 1 near the category median in speed and composite scoring."

Availability: How to Access Gemini Decision 1

Gemini Decision 1 is currently accessible through the Gemini API and Google AI Studio.

Authentication and SDK Configuration

Access requires a standard Gemini API key. Developers must configure their client libraries to point to the v1beta endpoint:

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})

response = client.decisions.create(
    model="gemini-decision-1",
    input=[
        {
            "type": "text", 
            "text": "User ticket #4081: Customer requests refund for unopened software license."
        }
    ],
    questions={
        "eligible_for_refund": {
            "type": "noul",
            "instructions": "Does the request meet the standard refund policy for unopened software?"
        },
        "ticket_priority": {
            "type": "choice",
            "instructions": "Determine the urgency of this support interaction.",
            "options": {
                "low": "Non-urgent, standard response window",
                "medium": "Needs response within 12 hours",
                "urgent": "Requires manager intervention"
            }
        }
    }
)

The model is distributed under standard Google developer terms. The free prototyping tier provides 200 requests per minute, 50,000 tokens per minute, and up to 10,000 requests per day. Pay-as-you-go tiers charge $0.10 per 1 million input tokens without rate limiting on standard production quotas. Free tier inputs may be used by Google to improve services, while paid commercial calls are excluded from model training.

While developers building generative reasoning pipelines still rely on models like Gemini 4 Carbon for multi-step tasks, Gemini Decision 1 is strictly restricted to classification primitives. Google Cloud Vertex AI enterprise support remains listed as "coming soon."

What We Don't Know Yet

As Gemini Decision 1 remains in early preview, several engineering questions remain unanswered:

1. Vertex AI Availability Timeline: Google has not announced a firm date for when enterprise customers can deploy the model within private Google Cloud VPC perimeters.
2. Context Window Latency Degradation: Latency metrics (150 to 250 milliseconds) were established on context windows under 16,000 tokens. Public reports have not documented latency curves when pushing inputs toward the 256,000-token maximum.
3. Multimodal Audio and Video Processing Limits: While native audio and video inputs are supported, Google has not published formal frame-rate sampling caps or per-minute tokenization equivalents specific to the Decisions API.
4. Third-Party Calibration Audits: Independent testing has yet to verify whether Google's reported 0.0250 ECE calibration score holds up outside curated classification suites.

Frequently Asked Questions

What is Gemini Decision 1?

Gemini Decision 1 is a multimodal decision model developed by Google DeepMind that evaluates bounded, closed-ended questions across a shared input in a single forward pass without generating text.

How much does Gemini Decision 1 cost to run?

Gemini Decision 1 costs $0.10 per 1 million input tokens, while output tokens are always free ($0) because the model does not decode generative text.

Is Gemini Decision 1 open source?

Gemini Decision 1 is not open source; it is a proprietary model accessible via Google AI Studio and the Gemini Decisions API.

How does Gemini Decision 1 differ from generative Gemini models?

Generative Gemini models decode freeform text token by token, whereas Gemini Decision 1 processes candidate choices simultaneously in a single pass to return fixed probabilities and confidence scores.

What input modalities can Gemini Decision 1 evaluate?

Gemini Decision 1 natively supports text, images, audio, video, and PDF documents within a single context window of up to 256,000 tokens.

How does Gemini Decision 1 compare to TypeSafe Jev?

In tests reported by Google, Gemini Decision 1 outperformed Jev 1.14 on 5 of 7 shared evaluations, scoring 92.2% versus 90.0% on JevBench, though its $0.10 per million input token price is higher than Jev's reported $0.042 rate.

---

What to watch next

Builders should track three signals over the coming weeks: the launch of enterprise VPC endpoints on Google Cloud Vertex AI, independent benchmark updates measuring calibration under maximum 256K-token loads, and upcoming model revisions such as the rumored Gemini 4 Hydrogen preview.

Try related models on Kie.ai: Gemini 4 Carbon, Gemini 4 Argon and Gemini 3.8 Flash.