Google Releases Multimodal Decision Model Gemini Decision 1 and Decisions API
Google has launched the Gemini Decisions API, powered by the multimodal decision model Gemini Decision 1, designed specifically for closed-ended multiple-choice questions, supporting three decision modes: Yes/No, multiple choice with up to 255 options, and ordered scoring. The model charges $0.10 per million input tokens, with output tokens free of charge; a single request can process 256K tokens of multimodal input and pose 1,000 questions simultaneously, with a median latency of 50-100 milliseconds at a 16K context. Google is also offering a free API tier in AI Studio, and integrations with OpenRouter, LangChain, Vercel, and others are in progress.
Why read this
SelectedAI score 79
The original text provides pricing, latency, and three decision modes, which can be used to assess the division of labor of decision models in agent routing.
Topics
Body · AI translation
After waiting for ages, Gemini 4 Hydrogen never showed up.
What Google served up last night was the multimodal decision model Gemini Decision 1.
Sundar announced: Today, Google officially launches the Gemini Decisions API, powered by the Gemini Decision 1 model.
Stitch Rapid uses the Decisions API to design an entire product flow in seconds
The "Jev" wind has finally blown over to Google's side.
It was born specifically for "closed-ended multiple-choice questions," with ultra-low latency, native multimodality, and the ability to handle 1,000 questions at once.
It has 3 native decision types: "Yes/No," "Multiple choice (supporting up to 255 options!)," and "rubric-based scoring."
It costs only $0.10 per million input tokens, and output tokens are completely free! (Just 1/5 of Flash-Lite.)
Across multiple benchmarks, it holds an absolute lead with the lowest calibration error.
Overturning decision models with multimodality
TypeSafe AI's Jev set off a wave of excitement in Silicon Valley, and other giants quickly followed suit. Google has finally entered the decision-model arena.
Since Jev has already driven the price of pure-text decisions down to $0.042 per million tokens, how should Google respond?
The answer: use multimodality to crush pure-text models.
You can feed it a massive amount of data—up to 256K tokens (whether text, images, audio, video, or PDFs)—all at once, and then fire 1,000 soul-searching questions at it simultaneously!
Whether you're doing tool calling, policy risk control, dynamic model routing, or context scoring, the Decisions API can give you the most decisive answer directly.
Gemini Decision 1 scores all options instantly through a single efficient "forward pass."
Topping the benchmarks: "intuition" that's ultra-precise
Of course, a decision model needs not only speed but also accuracy!
In top-tier benchmarks covering agent routing, logical reasoning, information ranking, and multimodal understanding, Gemini Decision 1 tops the charts across the board.
More importantly, it has the lowest calibration error across the board!
So however confident it is, that's how accurate it actually is, and its hallucination rate is very low.
Because it doesn't need word-by-word decoding, when Gemini Decision 1 processes multimodal contexts of up to 16K tokens, the model's median compute latency is kept at 50-100 milliseconds, and end-to-end latency is only 150-250 milliseconds.
Moreover, whether you ask 1 question or 100 questions at once, the speed is almost exactly the same—no lag at all.
Because it is nearly 5 times cheaper than the currently recognized "king of cost-effectiveness," Gemini 3.5 Flash-Lite (using the interactive API model), this is an absolute price butcher.
Native multimodality + three major decision modes: letting the model go all the way in one shot!
Built on Gemini's powerful native multimodal architecture, whether it's documents, recordings, or videos, it can directly "see and understand" them in a single forward pass.
You can mix and match various modalities—for example, pairing "customer service call recordings" with "account data," or bundling "PDF claim forms" with "damage photos" for it.
You can even freely adjust audio and video resolution to control visual detail and token costs.
In a single request, you can mix the three core decision modes.
Moreover, it thoughtfully returns a "confidence score (0 to 1)" every time; when things are ambiguous, your code can directly hand it off to a fallback or human review—very robust.
noul (yes/no): It determines the probability that a yes/no condition holds, and you can also choose a "true/false" description criterion.
choice (multiple choice): Supports 2 to 255 options! It directly tells you who the winner is, along with the probability ranking of all candidates.
score (precise scoring, ordered rubric): It scores the input content according to the rules you set (starting from 0), giving a continuous weighted score and the probability of each level.
from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
model="gemini-decision-1",
# Shared input: throw it an image + a sentenceinput=[{
"type": "image",
"mime_type": "image/jpeg",
"text": "Check the lounge bar counter before the 9 a.m. client meeting.",
"instructions": "Is there any uncleaned liquid spill on the bar counter?",
"instructions": "Which cleanup measure is most suitable for the current situation?",
"criteria": {
"wipe_with_cloth": "A small spill on the bar counter; a cloth is already nearby.",
"dispatch_janitorial": "Large hazardous area on the floor, broken glass, or continuous water leakage.",
"ready_for_meeting": "The bar counter is clean and ready.",
"instructions": "Assess the severity of the mess at the bar counter.",
Now, top teams inside Google have long treated it as a "low-latency control hub."
No more wasting time and effort training custom classifiers, and no more waiting for large models to slowly generate—just define the options, and the following features go live immediately!
For example, it can be dynamic model routing in Gemini Notebook.
In Notebook, the Decisions API can serve as a "traffic cop." As soon as a user's question comes in, it can classify the intent in a few milliseconds.
Simple questions go to a lightweight model (latency drops by 40%!), while complex questions are thrown to a deep-thinking model.
It can enable lightning-fast design iteration in Stitch: this is absolutely a job-killer of an application!
Stitch's fast canvas can generate a "runnable wireframe" of an entire product flow in just a few seconds.
Changes you make by voice or click take effect instantly. When the flow reaches the ideal state, you can switch to full design mode with one click.
You can also use it to instantly pick emojis in CC. In Google Labs' CC social app, it can select the most suitable one from more than 200 emojis in under a second. The emoji pops up instantly—very smooth.
It is also the security layer for MCP tools in Search. Google Search uses the Decision API to detect adversarial behavior and adversarially evaluated content with high efficiency.
Free API is here
And this time, Google AI Studio has also prepared a free API tier for us to freely use for prototyping.
Its billing logic is extremely fair: for shared input, no matter how many questions you ask, you are charged only once! (Each question is charged only for its own instructions and standard tokens.)
There are no output fees, no erratic "thinking token" assassins, and no separate audio transcription or OCR fees.
Not only that, Google's entire AI ecosystem is fully following suit! Integrations for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode, and the google-genai SDK are all being actively rolled out.
A battle of the gods: within 17 days, the giants frantically followed suit
This "decision model" storm triggered by Jev has completely torn up the old billing rules for large models—a single forward pass directly produces the result, costs plummet, and output is entirely free.
The whole of October can be called a "crazy month" for AI decision models. In just 17 days, multiple top AI labs released similar models.
Amazon: directly open-sourced a decision model called Strands Decider 2B.
Cloudflare: launched the Clef and Clef-flash models, open-sourced under the Apache 2.0 license.
Perplexity: opened a Decisions API fine-tuned from a 27B open-source model.
Liquid AI: launched d1-ultra.
Giants such as Microsoft and OpenAI also quickly adjusted at the API level, trying to block this wave.
In the fierce 17-day battle among the giants, Google's approach with Gemini Decision 1 is: don't compete on pure-text price differences, but go directly for "native multimodality + massive concurrency of a thousand questions."
The current competitive landscape is: for extreme low prices on pure text, choose Jev; for complex multimodality and high concurrency, look to Google; and for ecosystem compliance, leave it to OpenAI.
The future Agent architecture is already clear: the decision model (System 1) is responsible for making judgments at extreme speed and low cost, while the large model (System 2) retreats to the second line to handle deep "slow thinking."
The second half of the agent boom is here.
References:
This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelation, published by 36Kr with authorization.