Google launches Gemini Decisions API at $0.10 per 1M input tokens, free output

5 minutes ago

8 min read

Google has launched the Gemini Decisions API, a new Gemini API endpoint built for fast, closed-set decisions, together with a dedicated model called Gemini Decision 1 (model ID gemini-decision-1). The service was announced on November 11, 2026 in a post by Min Choi, product manager at Google DeepMind, and Felix Yu, principal scientist at Google DeepMind, and it is available now through the Gemini API and Google AI Studio.

The pricing is unusual for a Gemini model. Google charges $0.10 per 1M input tokens and $0 for output tokens, with no separate charge for thinking tokens, audio transcription or OCR. The shared input of a request is billed once, however many questions are asked about it, and each extra question adds only the tokens of its own instructions and answer criteria. A no-cost tier is available in Google AI Studio for prototyping, while support in Gemini in Google Cloud is described as coming soon.

The model does not write answers token by token. It scores a fixed set of candidate answers in a single forward pass and returns probabilities plus a confidence score between 0 and 1 for each question. One request can carry a single multimodal input of up to 256K tokens (text, images, audio, video or PDF) and up to 1,000 independent questions. Google quotes an end-to-end latency of 150 to 250 milliseconds, with a median model compute time of 50 to 100 milliseconds, for inputs up to 16K tokens, and says latency stays nearly flat between 1 and 100 questions per request.

··········

HOW THE GEMINI DECISIONS API AND GEMINI DECISION 1 WORK.

The Decisions API is aimed at the small, repetitive judgments that sit inside larger AI systems: choosing which tool an agent should call, checking a request against a policy, routing a query to a cheaper or stronger model, or grading whether retrieved context is relevant. Developers usually handle these steps today by prompting a general-purpose model and parsing its text output, which costs output tokens and adds generation time to every step. Gemini Decision 1 replaces that pattern with a scoring call that returns structured numbers directly.

Requests go to a new v1beta endpoint, POST https://generativelanguage.googleapis.com/v1beta/decisions, authenticated with a Gemini API key. A request has three required fields: the model name, an input (a string, a content object or a list of content objects, using the same content format as the Interactions API), and a map of questions. Each question has a type, text instructions and, for two of the three types, a list of criteria. Question keys may use letters, digits, hyphens and underscores, up to 64 characters.

The documentation defines three question types. A noul question is a yes/no check and returns the probability of yes, which the model always keeps between 0.01 and 0.99. A choice question picks one option from a fixed list of 2 to 255 entries and returns the winner with a probability for every option, and those probabilities always sum to 1 even when no option fits well. A score question grades the input against an ordered rubric whose levels start at 0, returning a probability-weighted score, the per-level probabilities and a legend. Google allows 2 to 255 rubric levels and recommends 2 to 10.

Every answer carries a confidence value, defined in the documentation as the estimated probability that the top answer is correct; for a yes/no question it equals the larger of the yes and no probabilities. Google presents this field as the hook for production logic, so that applications can send ambiguous cases to a fallback model or a human reviewer instead of acting on a low-confidence answer. The response also includes a usage object with input token counts by modality, and its total output token count is always zero.

........

Gemini Decisions API limit	Value
Model ID	gemini-decision-1
Questions per request	1 to 1,000
Options per choice question	2 to 255
Rubric levels per score question	2 to 255 (2 to 10 recommended)
Context window	262,144 tokens (input plus all questions)
Inline input size	Up to 100 MB, 50 MB for PDFs
External HTTPS file URLs	Up to 15 MB each
Latency claimed by Google	150-250 ms end to end, up to 16K tokens
Output tokens billed	None

........

The official quickstart shows the shape of a minimal call with the google-genai Python SDK, version 3.2 or later:

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
    model="gemini-decision-1",
    input="The blue mug I ordered on Monday arrived cracked on Wednesday.",
    questions={
        "is_damaged": {"type": "noul", "instructions": "Did the item arrive damaged?"},
    },
)
answer = result.answers["is_damaged"]
print(f"P(yes)={answer.noul:.2f} confidence={answer.confidence:.2f}")

The JavaScript SDK (@google/genai) and plain REST are also supported. Inputs can be passed inline as base64, through Files API URIs or through public or signed HTTPS URLs. YouTube links and gs:// Cloud Storage URIs are not accepted, and question instructions must be text.

··········

GEMINI DECISION 1 PRICING COMPARED WITH GEMINI 3.5 FLASH-LITE AND OPENAI'S DECISIONS API.

Gemini Decision 1 is a new model with no earlier version, so the useful comparison is with the two alternatives a developer would weigh today: running the same checks through Google's cheapest current text model, Gemini 3.5 Flash-Lite, or using OpenAI's competing Decisions API, which runs on GPT-6 Luna. The Gemini Decision 1 figures come from Google's launch announcement and the Decisions API documentation; when Data Studios checked Google's Gemini API pricing page on November 11, 2026, the model did not yet have its own entry there. The Gemini 3.5 Flash-Lite figures come from that pricing page, and the OpenAI figures from OpenAI's Decisions guide and GPT-6 Luna model page, all as listed on November 11, 2026.

........

Model (USD per 1M tokens)	Input	Output	Cached input	Batch	Effective date
Gemini Decision 1 (Decisions API)	$0.10	$0	Not supported	Not supported	November 11, 2026
Gemini 3.5 Flash-Lite (paid tier)	$0.30	$2.50	$0.03 + $1.00/1M tokens per hour storage	$0.15 input, $1.25 output	Already in effect
GPT-6 Luna via OpenAI Decisions API	$0.10 uncached	$0	Discounted (Luna lists $0.01)	Not listed for Decisions	Already in effect

........

Two limitations in the Gemini documentation shape the real cost. Context caching, the Batch API and the Flex and Priority inference tiers are all unsupported for Gemini Decision 1, and repeated input sent across separate requests is billed in full every time. OpenAI, by contrast, states in its Decisions guide that cache hits on a reused prefix of at least 1,024 tokens are discounted, while cache writes are charged and regional processing premiums and long-context multipliers can apply. The list prices for uncached input are identical at $0.10, so the gap between the two services depends on how much of each request repeats.

The capacity limits differ too. OpenAI's guide allows 1 to 400 questions per request and 2 to 10 score levels, while Google allows up to 1,000 questions and up to 255 rubric levels. OpenAI's guide also lists file IDs, files and audio among the unsupported inputs, whereas Gemini Decision 1 accepts audio, video and PDFs directly. Developers who followed our coverage of OpenAI's Decisions API launch with GPT-6 Luna will recognize the same basic design, with Google betting on larger requests and native multimodal input.

··········

WHAT A SUPPORT-TICKET ROUTER COSTS ON GEMINI DECISION 1 VERSUS GEMINI 3.5 FLASH-LITE.

A typical workload for this kind of model is an automated triage layer in front of a customer-support queue. The assumptions for this example are 50,000 tickets per day, a shared input of 1,500 tokens per ticket (the message plus account context), and five questions per ticket (is it a refund request, which of several queues it belongs to, how urgent it is on a rubric, whether it contains abusive content, and whether a human must review it), each adding about 60 tokens of instructions and criteria. For the generative baseline, the same five questions are asked in a single Gemini 3.5 Flash-Lite prompt with about 200 extra tokens of formatting instructions and a 150-token JSON answer, and thinking tokens are left out, which favors the baseline.

........

Ticket triage, 50,000 requests per day	Tokens per request	Daily cost	30-day cost
Gemini Decision 1, 5 questions	1,800 input, 0 output	$9.00	$270.00
Gemini Decision 1, 20 questions	2,700 input, 0 output	$13.50	$405.00
Gemini 3.5 Flash-Lite, standard	2,000 input, 150 output	$48.75	$1,462.50
Gemini 3.5 Flash-Lite, Batch API	2,000 input, 150 output	$24.38	$731.25

........

The arithmetic for the Decisions API is 1,800 tokens times 50,000 requests, or 90 million input tokens a day, at $0.10 per million, for $9.00 a day. The Flash-Lite baseline costs $30.00 a day for 100 million input tokens at $0.30 per million plus $18.75 for 7.5 million output tokens at $2.50 per million, for a total of $48.75, about 5.4 times the Decisions API figure. That ratio is close to Google's own claim that Gemini Decision 1 is nearly 5 times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across its benchmark suite. Quadrupling the questions to 20 per ticket raises the Decisions API bill by only half, to $13.50 a day, because the 1,500-token ticket is billed once.

Batch processing narrows the gap for workloads that can wait, since Flash-Lite at batch rates comes to $24.38 a day in this example, still well above the real-time Decisions API figure. The pricing structure clearly favors high-volume, latency-sensitive classification with many questions per input, such as agent tool routing, moderation and support triage. It helps less when the same long document is evaluated in many separate requests, because the absence of context caching means every request pays for the full input again.

··········

WHO CAN USE THE GEMINI DECISIONS API AND WHERE IT ALREADY RUNS.

Any developer with a Gemini API key from Google AI Studio can call the endpoint today on the v1beta API, and Google points to a no-cost tier in AI Studio for prototyping without publishing its limits in the announcement. The documentation links to rate-limit and pricing anchors for the model, but neither carried figures when checked on launch day. Readers planning production volumes can check our guide to Gemini API pricing in 2026 and our explainer on Google AI Studio rate limits and spend caps, which cover the tier system that applies to Gemini API projects. Enterprise teams on Google Cloud cannot use the model through Vertex AI yet, since Google lists that support as coming soon and offers a sign-up form.

Google says several of its own products already run on the API. Gemini Notebook uses it to route queries to the right model tier and thinking level, which Google says serves everyday queries with up to 40% lower latency. Google Search uses it as a safety layer to detect adversarial behavior in MCP tools, the Stitch rapid canvas experiment uses it to generate playable wireframes and apply voice or tap edits, and a Google Labs project called CC uses it to pick an emoji reaction from more than 200 options. Outside Google, the launch names OpenRouter, LangChain, the Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode as supported integrations alongside the google-genai SDKs.

For teams moving an existing classifier onto the new endpoint, the practical work is limited. Prompts written as instructions for a generative model need to be restated as typed questions with explicit options or rubric levels, the confidence field needs a threshold below which a case goes to review, and any workflow that depends on cached prefixes or batch jobs should be costed again under full input billing.

··········

GEMINI DECISION 1: CONFIRMED FACTS AND OPEN QUESTIONS.

The confirmed facts are the price of $0.10 per 1M input tokens with free output, the limits of 1,000 questions and 262,144 tokens per request, the three question types with calibrated probabilities and confidence scores, multimodal input including audio, video and PDFs, and availability through the Gemini Developer API and Google AI Studio from November 11, 2026.

Several points remain unverified or undisclosed. Google says Gemini Decision 1 leads its benchmark suite for agent routing, reasoning, ranking and multimodal understanding with the lowest calibration error, but the announcement text does not publish the scores or name every competing system, so the accuracy claim cannot yet be checked independently. The latency figures are Google's own measurements for inputs up to 16K tokens and say nothing about behavior near the 256K ceiling. The free-tier quota, paid rate limits and the date of Google Cloud availability are also unpublished. For developers choosing between Google and OpenAI, the choice depends on request shape, since Gemini Decision 1 offers larger question sets and native media input at the same $0.10 list price, while OpenAI's prompt-caching discount can make GPT-6 Luna cheaper for repeated long prefixes.

··········

·····

DATA STUDIOS

·····

datastudios.org