Gemini Decisions API: What SMBs Should Do Now vs Later

Gemini Decisions API: What SMBs Should Do Now vs Later: focused woman working on laptop in shop
Model decision routing: Gemini Decisions API: What SMBs Should Do Now vs Later

In short

Google's Gemini Decisions API scores fixed options like yes/no or category choices and returns a confidence number instead of writing text. For a small business, it's worth a low-risk prototype on routing or document checks now. Production dependence can wait until the integrations mature.

It’s worth testing, but as a prototype rather than a rebuild. On November 11, 2026, Google introduced the Gemini Decisions API, which doesn’t write answers. It scores a fixed set of options and tells you how confident it is. If your team already sorts requests, checks documents against rules, or grades content, that’s the work this is aimed at.

What actually changed with the Gemini Decisions API?

Most LLM calls generate text one token at a time. Google’s announcement describes something different: a model called Gemini Decision 1 that rates each of your candidate options in one pass through the model and returns probabilities. You send one input (text, images, audio, video or PDFs, up to 256K tokens) plus up to 1,000 independent questions.

There are three question types. Yes/no estimates the probability that a condition holds. Choice picks from 2 to 255 options and gives a probability for each. Score grades against an ordered rubric. Every answer carries a confidence score from 0 to 1, and Google says the point is to let your code send shaky cases to a fallback or a human reviewer.

Google quotes 150-250 ms end-to-end latency and a price of $0.10 per 1M input tokens, with no output tokens billed. It also says there are no separate charges for audio transcription or OCR. The shared input is billed once per request, however many questions you attach.

Key figuresGemini Decisions API at a glance, per Google
  • 150-250 msEnd-to-end latencyGoogle AI
  • $0.10 per 1MInput tokens, with zero output tokens billedGoogle AI
  • 1,000Independent questions per requestGoogle AI
  • 256K tokensMaximum multimodal inputGoogle AI

Sources: Google AI, Hugging Face, AWS Machine Learning

Why does this matter for a small or midsize business?

Because a lot of business AI work isn’t writing. It’s deciding: is this email a complaint or a quote request, does this contract page mention a termination clause, does this photo meet the checklist. Those are bounded choices, and until now you’d either train a classifier or pay for a full generative call to get one.

Focused developer at dual monitor setup

Google’s own example pairs a scanned PDF claim with damage photos, or a call recording with account metadata. Picture an imagined case: a distributor receives order emails with attached PDFs. One request could ask whether the order is complete, which department should handle it, and how urgent it is on a rubric. Anything below your confidence threshold goes to a person.

On cost, Google’s own comparison puts Gemini Decision 1 on the Decisions API at nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API, measured over the benchmarks Google chose. Those are Google’s tests, not yours. Treat the figure as a reason to run your own test, not as a number to plan a budget around.

Do gateways and local models change the picture?

Two other releases from the same week add context. AWS published a pattern for a self-service LLM gateway that puts LiteLLM in front of Amazon Bedrock, with per-user API keys, per-user cost controls and performance monitoring. It’s written for coding assistants, but the idea carries over: one OpenAI-compatible endpoint your tools call, with usage tracking behind it. It’s also not light. The AWS design involves Fargate, a PostgreSQL database and a cache.

LiteLLM is also on Google’s list of ecosystem integrations that are rolling out for the Decisions API, alongside OpenRouter, LangChain and Vercel. So a gateway layer may be how you try this without wiring a new SDK into every app.

On the privacy side, llama.cpp reached v1.0.0. The project says it supports more than 150 model architecture variants and will now flag breaking API and ABI changes with major-version bumps. That matters if you want to run models on your own hardware, since a stable interface makes upgrades less of a gamble. The page also links to a post titled “New in llama.cpp: Decision Models,” but doesn’t explain it, so read that before assuming how it relates.

What should you do now, and what can wait?

Do now: list the decisions your team makes repeatedly that have a small set of possible answers. Pick one where a wrong answer is cheap to catch. Google offers a no-cost tier in Google AI Studio for prototyping, so a first test doesn’t need a budget line.

How it worksHow a decision call works

Decisions API request

    1. Send one shared input plus your questions
    2. Model scores the options in one pass
    3. Probabilities and 0-1 confidence come back
    4. Is confidence high enough?
    No: send to fallback or human
  1. Act automatically

Source: Google AI

Wait on: anything that makes this service a hard dependency. Google says support in Gemini in Google Cloud is coming soon and several integrations are still rolling out. The model is new, and the performance claims come from Google. Hold off on sending sensitive customer data until you’ve read the data terms that apply to you.

How would you evaluate it as part of an LLM integration?

Start with a sample of past cases where you already know the right answer. Run them through and compare. The number to watch isn’t just accuracy but calibration: when the model says 0.9, is it right about nine times in ten? Google says Gemini Decision 1 has the lowest calibration error on its benchmarks, so check that against your own data.

Man sitting at desk discussing data on a laptop screen

Then set thresholds. Decide what confidence is high enough to act automatically, what goes to a person, and who that person is. Keep the model call behind a thin layer in your own code, or a gateway, so you can swap models without rewriting the workflow. Compare against what you do today, whether that’s manual sorting or a generative call, on speed, cost and error rate.

OptionBest forWatch out for
Decisions API (Gemini Decision 1)Fixed-choice calls such as routing, policy checks and rubric grading. Each answer comes with a confidence score.It’s a new service. The performance figures are Google’s own. Some integrations are still rolling out.
Generative model callOpen-ended output such as drafting or summarizingOutput has to be parsed and checked; slower than a scoring call
LLM gateway (LiteLLM on Amazon Bedrock)One endpoint, per-user keys and cost tracking across toolsExtra infrastructure to build and run
Local models (llama.cpp)Keeping inference on your own hardwareYou supply and maintain the hardware
Four ways to handle model work. Each one fits a different job.

What should you do next?

Pick one repetitive decision in your business and write down the possible answers and what a wrong one costs. Pull 50 to 100 past cases with known outcomes. Then run them through the free tier in Google AI Studio and see where the confidence scores line up with reality. That’s a few days of work, and it tells you whether this belongs in your LLM integration plans. Want help with this? See our work on LLM integration and AI agent development, or book a meeting with our team.

Frequently asked questions

Does the Decisions API replace a chatbot or a text-generating model?

No. It’s built for closed-set decisions where you define the options, and it returns probabilities rather than written answers. You’d still use a generative model for drafting, summarizing or conversation.

What does the Decisions API cost?

Google lists $0.10 per 1M input tokens and no charge for output tokens, plus a no-cost tier in Google AI Studio for prototyping. The shared input is billed once per request, and each extra question adds just the tokens for its own instruction and criteria.

Why does a confidence score matter?

It lets your code decide what to automate. High-confidence cases go straight through, and ambiguous ones go to a fallback or a human reviewer, which keeps mistakes from reaching customers unnoticed.

Sources

About the author

Aranv Editorial Team

Aranv's editorial team covers practical developments in AI and business software for small and midsize businesses. Each article explains what a change means for owners and operations teams, with links to the original sources.

Share on

Talk to our team

Adding AI to software you already use?

Tell us what you are trying to improve. We will reply with a useful next step, whether that is a short call, a scoped pilot or a pointer to the right approach.

Talk to an LLM expert How we help with LLM integration

Or send a short inquiry

    Add your timeline and tools (optional)

    Please do not include passwords or confidential records. See our Privacy Policy.

    Related articles

    What application do you have in mind?
    We are hands-on to deliver beyond your expectations

    Talk to an LLM expert



    Get in Touch

    5000 Centregreen Way, Suite 500
    Cary, NC 27513
    5820 N Canton Center Rd, Suite 170
    Canton, MI 48187