Gemini Decision 1: Google ships a decision model at $0.10 per 1M tokens

Published: November 12, 2026 · Official

On November 11, 2026 Google introduced the Gemini Decisions API, a separate endpoint powered by the new Gemini Decision 1 model. It does not write text: it scores up to 1,000 preset questions against one shared input and returns probabilities in 150–250 ms. Pricing is $0.10 per 1M input tokens with no output-token charge.

What Google announced

The announcement appeared on the Google blog on November 11, 2026, written by Min Choi, product manager at Google DeepMind, and Felix Yu, principal scientist. Gemini Decisions API is a new Gemini API endpoint running Gemini Decision 1, model ID gemini-decision-1. Instead of generating prose, the model scores candidate options and returns probabilities with confidence values.

The target is a closed set of decisions: tool selection, policy checks, routing queries between model tiers and thinking levels, and context grading. Google calls these purpose-built primitives for bounded choices — the branching calls inside agents where teams previously trained custom classifiers or waited for a generative response and parsed it.

The endpoint is live on the Gemini Developer API in v1beta and in Google AI Studio, with a no-cost tier for prototyping; it requires google-genai SDK 3.2 or later. There are no open weights and no self-hosting option. Google frames Gemini Decision 1 as its first model of this kind, though the category is older: TypeSafe AI shipped Jev in mid-September and OpenAI launched its Decisions API in October.

One shared input, up to 1,000 questions

A request has three parts: a model ID, a shared input, and a map of questions. The shared input can hold text, images, audio, video and PDFs — up to 256K tokens (262,144) — and modalities can be mixed, for example a call recording paired with account metadata. Up to 1,000 independent questions can be scored against that input in one pass; questions are scored in parallel and cannot see each other, and the shared input is billed once per request.

There are three question types. noul is a yes/no check that returns a probability bounded between 0.01 and 0.99, optionally with descriptions of the true and false conditions. choice picks one of 2 to 255 options and returns normalized probabilities for every candidate. score grades an input across an ordered rubric starting at 0 and returns a weighted score plus per-level probabilities.

The design difference from an LLM is that Gemini Decision 1 does not decode an answer token by token. It scores options in a single forward pass, which Google measures at 50–100 ms of median model compute and 150–250 ms end to end. Those figures hold for inputs up to 16K tokens and stay nearly flat from 1 to 100 questions per request.

Every answer carries a confidence value from 0 to 1, so applications can route ambiguous cases to a fallback or a human reviewer. Prompt caching is not supported, and there are no separate thinking-token, transcription or OCR charges.

Price, benchmarks and calibration

Input costs $0.10 per 1M tokens and output tokens are not billed at all. Google compared the model with TypeSafe Jev 1.14, GPT-6 Luna in OpenAI Decisions and its own Gemini 3.5 Flash-Lite. According to Google's table, Gemini Decision 1 leads 7 of 9 accuracy benchmarks and posts the lowest calibration error across the suite.

Its best result is 92.2% on JevBench (public split). Its weakest is 58.7% Macro-F1 on Amazon ESCI. On When2Call it trails Jev 1.14, 79.9% versus 82.0%. On total price across the seven shared evaluations Jev is also cheaper — $0.68 against $0.88 — so Gemini Decision 1 is not the cheapest option in the group.

Calibration is where the gap is clearest: Top-1 ECE is 0.0250 for Gemini Decision 1 versus 0.0494 for Jev 1.14, 0.0721 for GPT-6 Luna and 0.2104 for Gemini 3.5 Flash-Lite. That matters when setting confidence thresholds for automation, since lower calibration error makes rules like 'below 0.9 goes to a human' more reliable.

Against Gemini 3.5 Flash-Lite, Google claims roughly five times lower cost across the shared benchmark suite: $0.88 versus $4.14. All of these numbers are self-reported by Google, and refusals or safety blocks were counted as incorrect answers.

• Leads 7 of 9 accuracy benchmarks in Google's table
• JevBench public split: 92.2%
• Amazon ESCI Macro-F1: 58.7%, the weakest score in the set
• When2Call: 79.9% versus 82.0% for Jev 1.14
• Top-1 ECE: 0.0250 versus 0.0494 for Jev 1.14
• $0.10 per 1M input tokens, zero output cost

Where Google already uses it

Google is not waiting for outside developers. Gemini Notebook uses the API as a real-time router, classifying intent in milliseconds and sending each query to the right model tier and thinking level, which Google says cuts latency for everyday queries by up to 40% while reserving deep reasoning models for complex tasks.

In Stitch, the rapid canvas experiment builds playable wireframes of a product flow in seconds and applies voice or tap edits almost instantly. The CC assistant in Google Labs picks an emoji reaction from more than 200 options, and Google Search uses the API as a safety layer to detect adversarial behavior involving Model Context Protocol tools.

Framework and SDK integrations — LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode — are rolling out gradually. Support in Gemini on Google Cloud is listed as coming soon.

What to use right now

Gemini Decision 1 is a narrow tool for a specific class of tasks, and it is not in the API Market catalog yet. The same practical jobs — routing a request, classifying, checking rules, grading content — can be handled today by fast models available through a single key, with payment in rubles and no VPN.

For those tasks, gemini-3.8-flash is a very fast Google model, gpt-6-luna is quick and cheap for bots and high-volume processing, and gpt-6.1-terra costs less than Sol while outperforming Luna. If you need long context and document work, the catalog also carries larger models. Keep an eye on the API Market model catalog as new models from providers arrive.

Models you can use right now

Connect in a minute

from openai import OpenAI

client = OpenAI(
    base_url="https://api-market1.shop/v1",
    api_key="sk-your-key",
)

reply = client.chat.completions.create(
    model="gemini-3.8-flash",
    messages=[{"role": "user", "content": "Hi! Tell me about yourself."}],
)
print(reply.choices[0].message.content)

FAQ

Sources

• Gemini Decisions API: Fast, calibrated multimodal decisions — Google, 2026-11-11
• Google Unveils Gemini Decision 1: Its First-Ever Decision-Making AI Model That Tops 7 Out of 9 Benchmarks With Up to 92.2% Accuracy - 36Kr — 36Kr, 2026-11-12
• Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens — marktechpost.com, 2026-11-11
• Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass — the-decoder.com, 2026-11-11
• Google enters decision model market with Gemini Decisions API — siliconangle.com, 2026-11-10
• Google’s Gemini Decisions API processes 1,000 decisions in 150–250 milliseconds — interestingengineering.com, 2026-11-11

More news

• Gemini 4 Hydrogen: what Google launched on November 12
• Gemini 3.7 Flash: Google bets on super-fast coding and half the price
• Gemini 4 Pro leaks promise big benchmarks and a low price
• Qwen 4: confirmed facts, the latest teaser and the release status of Alibaba's new model
• Gemini Omni 1.2 Flash: 15-second clips, edits in conversation, $0.07 per second

Русская версия