Gemini Decisions API: Google’s New Model Scores 1,000 Questions in One Call

TL;DR: Google has launched the Gemini Decisions API, a new endpoint powered by a model called Gemini Decision 1 (gemini-decision-1). Instead of writing text, it answers up to 1,000 yes/no, multiple-choice or rubric-score questions about one shared input in a single pass, and returns probabilities plus a confidence score for each. Google says a request takes 150-250 ms end to end and costs $0.10 per million input tokens, with no output tokens billed. It’s live now in Google AI Studio, including a no-cost tier for prototyping.

The Gemini Decisions API, announced on November 11, 2026 by Google DeepMind’s Min Choi and Felix Yu, targets a part of AI apps that rarely gets its own product: the small judgment calls. Which tool should an agent use? Does this message break a policy? Which model should handle this request? Most developers answer those today by prompting a chat model and parsing its text reply. Google’s pitch is that a model built only to score fixed choices can do that faster, cheaper and more reliably. Here’s what launched, how it works and who should care.

What is the Gemini Decisions API?

In Google’s announcement, the company describes it as “a new endpoint for fast, closed-set decisions.” You send one input, which can mix text, images, audio, video and PDFs up to 256K tokens, together with up to 1,000 independent questions. Rather than generating an answer token by token, Gemini Decision 1 scores the candidate options in a single forward pass.

Google lists four headline uses, in its own words “purpose-built primitives for bounded choices like tool selection, policy checks, model routing, and context grading.” Put simply, it’s built for the moments when your code needs a verdict to branch on, not a paragraph for a person to read.

Why “no text” matters

The Decisions API documentation makes the case directly: generative models produce text for people, and when code needs a judgment, “generating and parsing text adds latency and brittle parsing logic.” Because the API generates zero text tokens, output token usage is always zero, and your program gets typed values it can sort, route and compare without regex or JSON repair.

The three question types

Every question in a request uses one of three types. You can mix them freely against the same input.

The confidence score, from 0 to 1, is the practical hook. Google says it lets “your code route ambiguous cases to a fallback or human reviewer.” In other words, you can auto-approve the clear cases and send only the uncertain ones to a bigger model or a person.

Questions are isolated on purpose

According to the documentation, every question runs against the shared input in the same forward pass, and questions can’t see each other. Google’s recommended pattern is to ask narrow, single-purpose questions and then combine the answers in ordinary code with if statements and weights, rather than trying to cram a complex decision into one prompt.

Speed and price: Google’s numbers

All of the performance figures below are Google’s own claims from its launch post, not independent tests.

• Latency: 150-250 ms end to end, with median model compute of 50-100 ms for inputs up to 16K tokens. Google says latency stays “nearly flat” from 1 to 100 questions per request.
• Price: $0.10 per 1 million input tokens and $0 for output tokens.
• Shared input billed once: the main input is charged once per request no matter how many questions you ask; each question adds only its own instruction and criteria tokens.
• No extra fees: Google says there are no variable thinking-token bills or separate audio transcription and OCR charges.
• Cost comparison: Google claims the setup is “nearly 5x cheaper” across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.

For context on Google’s wider pricing push this month, see our coverage of Gemini 4 Carbon’s public release and pricing.

What could you actually build with it?

Google and early coverage point to the same kind of jobs: lots of small, repeatable judgments in front of more expensive work.

1. Agent tool selection: before an agent acts, ask which of its tools fits the request, and only call a large model when confidence is low.
2. Model routing: decide whether a query needs a cheap fast model or a premium one. Google says it already uses the API for a real-time router in Gemini Notebook and credits it with up to 40% lower latency on everyday queries.
3. Policy and safety checks: score content against moderation rules. Google says it also uses the API as a safety layer for MCP tools in Search.
4. Claims and support triage: VentureBeat describes a claims intake system sending a scanned form, damage photos and call notes in one request, then asking which queue it belongs in, whether evidence is missing and how urgent it is.
5. Grading and ranking: score drafts, search results or retrieved context against a rubric before you show them to a user.

If you’re building agents, these checks sit naturally alongside techniques like prompt chaining, and a fast policy check is one more layer of defense against prompt injection, though no classifier is a complete fix on its own.

Where it’s available

Google is also offering three demo apps, called Doorstep Parcels, Speed Sketch and Find the Zoo Keeper, to show the API in action.

How it fits the competition

Google isn’t alone here. VentureBeat notes that the launch puts Google into a market where OpenAI already has its own Decisions API running on GPT-6 Luna, Amazon’s Strands Decider lets developers run a model on their own infrastructure, and smaller players such as TypeSafe’s Jev handle low-cost classification. The common thread is that AI vendors now see routing and guardrail calls as a product category of their own, separate from chat.

What to watch before you switch

• Benchmarks are Google’s own. The speed and cost claims come from Google’s shared suite. Test it on your own routing or moderation prompts first.
• Closed-set only. It can only pick among options you define. Open-ended answers still need a generative model.
• Cloud availability. Teams standardized on Google Cloud’s enterprise Gemini offering will need to wait for the “coming soon” support.

The bottom line

The Gemini Decisions API is a small-looking launch with a big practical angle. By turning yes/no checks, multiple choice and rubric grading into a dedicated, cheap and fast endpoint, Google is going after the many hidden calls that make AI agents slow and expensive. If you build with AI, it’s worth benchmarking against your existing classifier or router prompts, starting with the cases where a confidence score could save you a call to a bigger model. And if you’re new to building with AI tools, our guide to vibe coding is a good place to start.

Sources

• Google: Gemini Decisions API announcement
• Google AI Studio: Decisions API documentation
• VentureBeat: Gemini Decision 1 scores up to 1,000 questions in one API call
• MarkTechPost: Google launches Gemini Decisions API

About Teck Hustlers Editorial Team

The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.