Google Unveils Its First Decision-Making AI Model: 1M Tokens Priced As Low As $0.1 With 100% Free Full Output Access

What is the internal culture and operation mode like inside the tech giant Google?

After waiting for a long time, Gemini 4 Hydrogen did not arrive.

What Google rolled out last night was the multimodal decision-making model Gemini Decision 1.

Sundar Pichai announced that Google has officially launched the Gemini Decisions API today, which is powered by the Gemini Decision 1 model.

Stitch Rapid uses the Decisions API to design the entire product workflow in seconds

The trend led by Jev has finally reached Google.

It is built exclusively for closed-ended multiple-choice questions, featuring ultra-low latency, native multimodality, and the capability to process 1000 questions at a time.

It has 3 native decision types: Yes/No, multiple choice (supports up to 255 options!) and scoring by predefined criteria.

It costs only $0.10 per million input tokens, with output tokens completely free! (Only 1/5 of the cost of Flash-Lite).

In multiple benchmark tests, it takes an absolute lead with the lowest calibration error.

Subverting Decision-Making Models With Multimodality

Jev from TypeSafe AI has set off a boom in Silicon Valley, and other tech giants have followed suit. Google has finally stepped into the decision-making model track.

Now that Jev has driven the price of plain text decision-making down to $0.042 per million tokens, how will Google break the deadlock?

The answer is: to completely outperform plain text models with multimodality.

You can feed massive data of up to 256K tokens (whether it is text, image, audio, video or PDF) to it at one go, and send 1000 critical questions to it simultaneously!

No matter you need to implement tool calling, policy risk control, dynamic model routing or context scoring, the Decisions API can directly give you the most straightforward answer.

Gemini Decision 1 instantly scores all options through one efficient forward pass.

Topping Benchmark Tests: Extremely Accurate "Intuition"

Of course, a decision-making model requires not only high speed, but also high accuracy!

In top-tier benchmark tests covering agent routing, logical reasoning, information sorting and multimodal understanding, Gemini Decision 1 tops the full leaderboard.

More importantly, it has the lowest calibration error across all tests!

That means the confidence level it outputs matches its actual accuracy, leading to an extremely low hallucination rate.

Since it does not need to decode word by word, when Gemini Decision 1 processes multimodal context of up to 16K tokens, the median model calculation latency is controlled at 50-100 milliseconds, and the end-to-end latency is only 150-250 milliseconds.

Moreover, no matter you ask 1 question or 100 questions at a time, the speed remains almost exactly the same, with no lag at all.

Since it is nearly 5 times cheaper than the currently recognized "cost-performance king" Gemini 3.5 Flash-Lite (interactive API model), it is an absolute price killer.

Native Multimodality + Three Decision Modes: One-stop Processing for the Model!

Based on Gemini's powerful native multimodal architecture, it can directly "read and understand" documents, recordings and videos in a single forward pass.

You can mix different types of multimodal content - for example, splice customer service call recordings with account data, or upload PDF claim forms together with damage photos.

You can even freely adjust the resolution of audio and video to control visual details and token costs.

You can use three core decision modes in combination in one single request.

In addition, it will thoughtfully return a "confidence score (from 0 to 1)" every time. In ambiguous scenarios, your code can directly redirect the task to a backup plan or manual review, which is very stable.

noul (Yes/No): It judges the probability of the yes/no condition being true, and supports selecting the description criteria for "True/False".

choice (Multiple Choice): It supports 2 to 255 options! It directly tells you the winning option, and provides the probability ranking of all candidates.

score (Precise Scoring, Ordered Scoring Criteria): It scores the input content according to the rules you set (starting from 0), and outputs continuous weighted scores and the probability of each level.

The interface is extremely user-friendly, using the familiar Content block format.

Let's look at the extremely elegant code below - how to make the AI look at a photo of an office bar counter and complete three different types of judgments in one single request.

from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
model="gemini-decision-1",
# Shared input: pass an image + a sentence to it input=[{
"type": "image",
"uri": "https://storage.googleapis.com/generativeai-downloads/images/spill.jpg",
"mime_type": "image/jpeg",
},
{
"type": "text",
"text": "Lounge bar counter check before the 9 a.m. client meeting.",
},
],
# Send 3 questions at one go!
questions={
"has_spill": {
"type": "noul",
"instructions": "Is there any uncleaned liquid spill on the bar counter?",
},
"next_action": {
"type": "choice",
"instructions": "Which cleanup measure is most suitable for the current scenario?",
"criteria": {
"wipe_with_cloth": "Small spill on the bar counter, rags are available nearby.",
"dispatch_janitorial": "Large-scale hazard on the ground, broken glass or continuous water leakage.",
"ready_for_meeting": "The bar counter is clean and ready for use.",
},
},
"severity": {
"type": "score",
"instructions": "Evaluate the severity of the mess on the bar counter.",
"criteria": [
"Clean and ready for use.",
"Slight local spill.",
"Major spill or safety hazard.",
],
},
},
)

How are top teams inside Google using it?

At present, top teams across Google have already treated it as a "low-latency control hub".

There is no need to spend time and effort training custom classifiers, no need to wait for large models to generate results slowly. Just define the options, and the following functions can go live directly!

For example, it can act as dynamic model routing in Gemini Notebook.

In Notebook, the Decisions API can be used as a "traffic police". Once a user's question comes in, it can classify the intent within a few milliseconds.

Simple questions are sent to lightweight models (latency reduced by 40%!), and complex questions are forwarded to deep thinking models.

It can achieve extremely fast design iteration in Stitch: This is an application that completely disrupts existing workflows!

The quick canvas of Stitch can generate a "runnable wireframe" for the entire product process in just a few seconds.

Modifications you make via voice input or clicks take effect instantly. When the process reaches the ideal state, you can switch to the full design mode with one click.

You can also use it to select emojis in CC in seconds. In the CC social app of Google Labs, it can select the most suitable emoji from more than 200 emojis in less than one second. The emoji pops up instantly, which is extremely smooth.

It also acts as the security layer for MCP tools in Search. Google Search uses the Decision API to detect adversarial behaviors and content for adversarial evaluation, with very high efficiency.

Free API is now available

Moreover, Google AI Studio has prepared a free API tier for us to prototype our projects for free.

Its billing logic is extremely user-friendly: no matter how many questions you ask for the shared input, you only pay once! (Each question is only charged for its own instructions and standard tokens).

There are no output fees, no unexpected "thinking token" hidden charges, and no separate fees for audio transcription or OCR.

Furthermore, Google's entire AI ecosystem is fully catching up! Integrations for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode and the google-genai SDK are all being rolled out rapidly.

Battle of Titans: Giants are Catching Up Wildly Within 17 Days

The "decision-making model" storm triggered by Jev has completely torn up the old billing rules for large models - results are generated directly in one single forward pass, the cost plummets, and all outputs are free of charge.

The whole October can be regarded as the "crazy month" for AI decision-making models. In just 17 days, many top AI labs have released similar models.

Amazon: It directly open-sourced a decision-making model named Strands Decider 2B.

Cloudflare: It launched the Clef and Clef-flash models, open-sourced under the Apache 2.0 license.

Perplexity: It opened the Decisions API fine-tuned based on a 27B open-source model.

Liquid AI: It launched d1-ultra.

Giants including Microsoft and OpenAI have also adjusted their API offerings rapidly, trying to block this wave.

In the 17-day fierce battle among giants, Google's strategy for Gemini Decision 1 is: instead of competing on the lowest price for plain text, it directly brings "native multimodality + concurrent processing of 1000 questions".

The current landscape of this track is: choose Jev for the lowest price in pure text scenarios, choose Google for complex multimodal and high-concurrency scenarios, and choose OpenAI for ecosystem compliance needs.

The future Agent architecture is already clear: the decision-making model (System 1) is responsible for making judgments quickly and at low cost, while large models (System 2) move to the second line and are responsible for deep "slow thinking".

The second half of the rapid development of intelligent agents has arrived.

References:

https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/

This article is from WeChat official account "AI Era", author: ASI Revelation, published with authorization from 36Kr.