Super Sale WeekClaude Skills — 20% OFF
News

Gemini Decisions API: What's New, Pricing, and Alternatives (2026)

Powerdrill Bloom·
Gemini Decisions API: What's New, Pricing, and Alternatives (2026)

Google launched the Gemini Decisions API on November 11, 2026. It is a new endpoint for fast, closed-set decisions, powered by a new model called Gemini Decision 1.

Instead of writing an answer, the model scores a fixed set of options. Google says it returns probabilities and confidence scores in 150 to 250 milliseconds, end to end. You can ask up to 1,000 questions about one input in a single request.

This guide covers what the API does, the three question types, pricing, and early uses inside Google. It ends with alternatives, including a no-code route for teams that want to score feedback in a spreadsheet. Details come from Google's launch post, read on November 12, 2026.

What Google announced on November 11

Google's post describes the product in one line. It calls the API "a new endpoint for fast, closed-set decisions powered by our new Gemini Decision 1 model."

The key facts from the post:

  • Model name. The model ID is gemini-decision-1.
  • Where it runs. It is available through the Gemini API in Google AI Studio.
  • Inputs. One input can mix text, images, audio, video, and PDFs, up to 256K tokens.
  • Questions. You can attach up to 1,000 independent questions to that input.
  • Speed. Answers arrive in 150 to 250 milliseconds end to end.
  • Price. Input costs $0.10 per 1M tokens, and output tokens cost nothing.

Google frames it as a building block for other AI systems. The post lists "tool selection, policy checks, model routing, and context grading" as target jobs.

How the Gemini Decisions API works

Most language models answer by generating text one token at a time. Gemini Decision 1 skips that step for bounded questions.

The post says the model "scores your candidate options in a single forward pass." It returns a probability for each option instead of a written reply. That design explains both the speed and the price.

Question design does most of the work. Google's example gives each choice a short description, such as "Small counter spill with a cloth or towel already nearby." Clear criteria like that give the model something concrete to score against.

A few habits help when you write your own questions:

  • Keep options separate. Two categories that overlap will split the probability between them.
  • Describe each option. A one-line definition beats a single word like "urgent."
  • Order rubric levels. Rubric scores start at 0, so put the lowest level first.
  • Add a catch-all. An "other" choice stops odd items from being forced into the wrong bucket.

Every answer carries a confidence score from 0 to 1. Google suggests using it to send unclear cases elsewhere. The post says code can "route ambiguous cases to a fallback or human reviewer."

Speed holds up as requests grow. Google says median latency stays "nearly flat from 1 to 100 questions per request," for inputs up to 16K tokens.

The three question types

The API supports three kinds of questions. You can mix all three against the same input in one request.

Type What it does Output
Yes/no (noul) Estimates the probability that a condition holds Probability that the answer is yes
Multiple choice (choice) Picks from 2 to 255 options The winner, plus normalized probabilities for every option
Rubric score (score) Grades input on an ordered scale starting at 0 A weighted score, plus a probability for each level

Google's own example checks a photo of a breakroom counter with a short text note. One request asks whether there is a spill, which cleanup action fits, and how severe the mess is.

For business data, the same pattern maps onto common review work. A support team could ask, for each ticket, whether it mentions a refund, which product area it concerns, and how urgent it is. Those are our examples, not Google's.

Multimodal input in one pass

The model reads several formats together. Google's post says it can process "text, images, audio, video, and PDFs directly in a single forward pass."

The post gives two pairings as examples. One is "a call recording with account metadata." The other is "a scanned PDF claim with damage photos."

That second example points at insurance and finance work. A claims team could score each claim file for completeness, pick a category, and rate damage severity in one call.

You can also tune image and video resolution. Google says this controls "visual detail and token cost."

How Google teams already use it

Google's post lists early uses inside the company:

  • Gemini Notebook. It uses the API to route each query to the right model tier and thinking level. Google says everyday queries get "up to 40% lower latency."
  • Stitch. A rapid canvas experiment uses it to generate playable wireframes of a product flow in seconds.
  • CC from Google Labs. It picks an emoji reaction from more than 200 options "in a fraction of a second."
  • Google Search. It acts as a safety layer that detects adversarial behavior in MCP tools.

Each case trades a long answer for a fast, bounded choice. None of them needs the model to write prose.

What it costs

Google's post gives a simple price. Gemini Decision 1 "is priced at $0.10 per 1M input tokens and $0 for output tokens."

There is also a free route for testing. The post mentions "an API no-cost tier you can use in Google AI Studio for prototyping."

Billing has a useful quirk for large question sets. Google says "Your shared input is billed once per request no matter how many questions you evaluate." Each extra question adds only its own instruction and criteria tokens.

Some common hidden costs are absent. Google says there are "no output-token charges, variable thinking-token bills, or separate audio transcription and OCR costs."

Item Price in Google's post
Input tokens $0.10 per 1M tokens
Output tokens $0
Prototyping in Google AI Studio No-cost tier

Here is a rough worked example, using our own assumptions. A team scores 5,000 survey answers, one request per answer. Each request carries about 400 tokens of answer text and 150 tokens of questions.

That is 550 tokens per request, or 2.75 million tokens in total. At $0.10 per 1M input tokens, the batch costs about $0.28. Real token counts depend on your text and your question wording, so test with a sample first.

Google also compares cost with its own generative models. The post says the Decisions API is "nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API."

Who can use it, and how to start

The Gemini Decisions API is a developer product. You need an API key from Google AI Studio and some code.

Google's post points to Python, JavaScript, and REST guides in the documentation. It also lists three demo apps: Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper.

Teams already on the Gemini API will find the input format familiar. Google says the input field "uses the same Content block format as the Interactions API." That means existing code that builds prompts with images or PDFs can be reused.

The output needs handling on your side. Your code reads the probabilities, applies a threshold, and decides what to do with low-confidence items. Plan that logic before you send the first large batch.

Third-party support is rolling out too. The post names OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. Support in Gemini in Google Cloud is listed as "coming soon."

Current limitations

A few limits follow from the post itself:

  • Closed questions only. The model picks from options you define. It does not write explanations or summaries.
  • Code required. There is no point-and-click interface described for business users.
  • Your options shape the answer. A poorly worded rubric will produce confident but unhelpful scores.
  • Google Cloud support is pending. Teams that buy through Google Cloud must wait for the listed integration.

Calibration still needs checking on your own data. Google reports the lowest calibration error on its benchmark suite, but your tickets or claims may differ.

What this means for teams that score data at scale

The Gemini Decisions API makes a common task cheap: putting labels on large piles of items. Support tickets, survey answers, product reviews, and claim files all fit.

Three questions help decide whether it fits a team:

  1. Are your questions closed? Yes/no checks, categories, and ratings suit it well. Open questions do not.
  2. Do you have a developer? Someone must write the calls and handle the confidence scores.
  3. Is speed or volume the problem? The API shines when you have thousands of items or need answers in under a second.

Teams with a few hundred rows a month and no developer may find it more work than it saves. For them, a no-code tool that reads the file and writes the summary is usually the shorter path.

Volume also changes the economics of review. When labels cost fractions of a cent, a team can score every ticket instead of a sample. The bottleneck then moves to acting on the results.

Alternatives to the Gemini Decisions API

Other ways to label and score data exist. Each has a different setup cost.

Option What it offers Price or access
Gemini 3.5 Flash-Lite with the Interactions API Generative model that can return labels as text Google's comparison baseline; Google says the Decisions API is nearly 5x cheaper on its benchmark suite
Custom-trained classifier A model trained on your own labeled examples Needs labeled data and machine learning skills
Powerdrill Bloom Themes, sentiment, and charts from uploaded feedback files, with no code Free plan at $0.00, Pro at $13.27/mo

Generative models. A general model can answer the same questions in text. It is more flexible, since it can also explain its choice. Google's post positions the Decisions API as the faster and cheaper option for bounded choices.

Custom classifiers. Google's post says its teams ship features "simply by defining options rather than training custom classifiers." A trained classifier still makes sense when you have years of labeled data and fixed categories.

Where Powerdrill Bloom fits

Many teams that need to score feedback do not write code. They have a CSV export from a survey tool or an app store, and they need a summary for a meeting.

Powerdrill Bloom handles that from the file itself. The AI Review Summary Generator groups reviews into "themes, sentiment and a leadership-ready callout." You upload a "CSV export from app stores, NPS, support tools or any review platform."

For customer interviews and open answers, the Voice of Customer Summarizer turns feedback into themes. You can then turn the results into charts with the AI Graph Maker.

The pricing page lists a Free plan at $0.00 and Pro at $13.27/mo.

Try Powerdrill Bloom with last month's survey export and ask for the top themes by sentiment.

Frequently asked questions

What is the Gemini Decisions API?

The Gemini Decisions API is a Google endpoint for fast, closed-set decisions. It scores options for yes/no, multiple choice, and rubric questions instead of writing text, and returns probabilities with confidence scores.

What is Gemini Decision 1?

Gemini Decision 1 is the model behind the Decisions API, with the ID gemini-decision-1. Google launched it on November 11, 2026, and it reads text, images, audio, video, and PDFs.

How much does the Gemini Decisions API cost?

Google prices Gemini Decision 1 at $0.10 per 1M input tokens and $0 for output tokens. There is also a no-cost tier in Google AI Studio for prototyping.

How many questions can one request include?

One request can include up to 1,000 independent questions about a single input. The input can be up to 256K tokens and can mix several formats.

Does the Gemini Decisions API write text answers?

No. It returns probabilities and confidence scores for options you define. For written summaries or explanations, you need a generative model or another tool.

Sources: Google, Gemini Decisions API: Fast, calibrated multimodal decisions.