Skip to content

Google’s Gemini Decisions API Answers 1,000 Questions in One Pass at $0.10 per Million Tokens, but Google’s Own Chart Shows a Cheaper Rival

Google's Gemini Decisions API scores 1,000 questions in one pass at $0.10 per million input tokens, but its own benchmark chart shows TypeSafe ahead on two tests and cheaper overall.

Gemini Decisions API official header image

Google released the Gemini Decisions API on 11 November 2026, a scoring endpoint that answers up to 1,000 questions about one shared input. It returns calibrated probabilities, not generated text, so code can branch without parsing a paragraph.

The model is Gemini Decision 1, available in the Gemini API and Google AI Studio. It costs $0.10 per million input tokens, bills nothing for output, and has a free tier of 200 requests a minute and 10,000 a day.

The catch is in Google’s own benchmark chart. Gemini Decision 1 wins seven of nine tests and has the lowest calibration error, but TypeSafe Jev 1.14 wins two tests and costs less on the shared suite. The “leads across the suite” claim depends on which columns you read.

Quick answer: Gemini Decision 1 is a fast classifier for closed choices, not a chatbot. It scores up to 1,000 questions over text, images, audio, video or PDFs, with a 262,144 token limit, at $0.10 per million input tokens. It suits routing, policy checks and grading, but it is not the cheapest or the best on every test.

What Google actually launched

The Decisions API is a new endpoint, not a chat model. You send one input and a map of questions, and each question returns a value plus a confidence score from 0 to 1.

The yes or no type, called noul, returns the probability that a statement holds. Choice picks one of 2 to 255 options and returns probabilities for each. Score grades an input against ordered rubric levels.

Questions are scored independently in one forward pass. They cannot see each other, so adding one does not change the others. That suits tool selection, routing, grading and policy checks.

The model does not generate text, call functions or run code. It also lacks the Batch API, context caching and structured outputs. The SDK needs google-genai 3.2 or later.

The limits that matter

The input limit is 262,144 tokens and one request can carry 1,000 questions. Inline data can reach 100MB, or 50MB for a PDF; a public HTTPS URL can be 15MB.

Text, images, video, audio and PDFs share one request. Video frames and PDF pages count as image tokens, so a long video can cost more than its text suggests.

The free tier allows 200 requests a minute, 50,000 tokens a minute and 10,000 requests a day. Tier 1 raises that to 1,500 requests a minute, 1 million tokens a minute and 50,000 a day.

There is no batch discount. Google says the API does not support the Batch API. Safety filters can reject a request with HTTP 400 instead of answers.

The price is the pitch

At $0.10 per million input tokens, Gemini Decision 1 is priced like a utility, not a frontier chatbot. Google bills the shared input once, however many questions you ask; each question adds only its instructions and criteria.

There are no output, thinking, transcription or OCR charges. That matters when the alternative is a chat model that transcribes a call and reads a scanned claim.

Google claims 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end, nearly flat from 1 to 100 questions. Those are vendor figures, with no independent test.

Google says its own products already use the API. Gemini Notebook routes queries and cuts latency by up to 40%, Stitch builds wireframes, and Google Search screens MCP tools for adversarial behavior. Those are internal claims, not independent results.

Google’s own chart has two cracks

The chart compares Gemini Decision 1 with TypeSafe Jev 1.14, GPT-6 Luna on the OpenAI Decisions API and Gemini 3.5 Flash-Lite as a general baseline.

Benchmark Gemini Decision 1 TypeSafe Jev 1.14 GPT-6 Luna Gemini 3.5 Flash-Lite
JevBench 92.2% 90.0% 90.9% 86.6%
Amazon ESCI 58.7% 56.7% 54.7% 47.2%
New Yorker Caption 79.7% 72.0% 68.4% 73.7%
Mind2Web Multimodal 86.2% 86.8% 81.8% 78.0%
PubMedQA 79.8% 78.2% 74.4% 77.6%
Global-MMLU-Lite 86.0% 85.0% 79.6% 78.8%
When2Call 79.9% 82.0% 70.4% 60.9%
MMAU 71.8% not listed not listed 67.7%
Video-MME 69.1% not listed not listed 63.3%
Calibration error, lower is better 0.0250 0.0494 0.0721 0.2104
Total price, seven shared evals $0.88 $0.68 $0.93 $4.14

Gemini Decision 1 wins JevBench, Amazon ESCI, New Yorker Caption, PubMedQA and Global-MMLU-Lite against both direct rivals. It loses Mind2Web Multimodal to TypeSafe, 86.8% to 86.2%, and When2Call, 82.0% to 79.9%.

Its two largest wins, MMAU and Video-MME, have no TypeSafe or GPT-6 Luna score. They beat only the Gemini 3.5 baseline, so the suite lead is narrower.

The five times price claim is also narrow. Gemini Decision 1 costs $0.88 on the shared suite against $4.14 for Gemini 3.5 Flash-Lite, but TypeSafe is cheaper at $0.68 and GPT-6 Luna is close at $0.93. The comparison is with Google’s older general model, not the specialists.

The calibration number is the strongest result. A top-1 error of 0.0250 means the confidence scores track observed accuracy better than the rivals, which matters when low-confidence cases go to a human.

Confirmed vs Not Confirmed

Claim Status
Gemini Decisions API launched on 11 November 2026 with the gemini-decision-1 model Confirmed by Google’s blog, model page and API docs
Up to 1,000 questions, 262,144 input tokens and zero output tokens per request Confirmed by the model page and API reference
$0.10 per million input tokens, free tier, no output charge Confirmed by Google’s pricing page
150 to 250 milliseconds end to end and nearly flat latency up to 100 questions Not independently verified; Google’s own figures
Leads across the benchmark suite Partly true; Google’s chart shows TypeSafe ahead on 2 of 9 tests and cheaper overall
Google Notebook cuts latency by up to 40% with the API Not independently verified; Google’s internal account

What it means for Indian developers

For Indian teams, the appeal is cost control on boring decisions. A startup that pays a chat model to classify tickets, check a policy or pick a tool can move those calls to a model that bills only input tokens. A million short checks cost ten cents.

The free tier suits a prototype, not production. At 200 requests a minute and 10,000 a day, a busy queue hits the ceiling quickly. Budget for Tier 1 and watch per-question tokens, because 1,000 questions on a large PDF is not free just because the input is billed once.

Data handling needs a separate look. The free tier is used to improve Google’s products and the paid tier is not, according to the pricing page. India’s DPDP Act duties follow personal data into the model call, and a confidence score does not create a lawful basis.

The API is also a lock-in decision. Prompts and thresholds are portable, but the calibration curve is not. Keep a labelled sample of your own traffic and test any switch instead of trusting a public benchmark.

The wider signal is architectural. Google, OpenAI and Microsoft now sell a cheap decision layer beside the expensive generation layer. Use a small scorer for the branch and save the large model for the answer. The OpenAI Decisions API and Microsoft-Decision-1 arrived in October. Google is the third entrant.

Competition is good for buyers, and the price should be read as an opening offer. Microsoft launched at $0.042 per million input tokens, less than half Google’s price. Google’s edge is multimodal input and calibration, not the lowest sticker price.

FAQ

What is the Gemini Decisions API?

It is a Google API that scores closed-set questions against one shared input and returns calibrated probabilities, with no generated text.

How many questions can one request ask?

Up to 1,000 independent questions, over an input of up to 262,144 tokens. They are scored separately.

How much does Gemini Decision 1 cost?

$0.10 per million input tokens on the paid tier, with no output charge. The free tier trains Google’s products.

Can it read images, audio and PDFs?

Yes. It accepts text, images, video, audio and PDF documents in one request, with resolution controls for images and video.

Is it better than TypeSafe Jev or GPT-6 Luna?

It depends on the task. Google’s chart shows Gemini Decision 1 ahead on most tests and best calibrated, while TypeSafe wins Mind2Web and When2Call and costs less across the shared suite.

Sources

Share this article

Leave a Reply

Your email address will not be published. Required fields are marked *

Loading the next article…

Continue reading