Google has launched the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). The Gemini Decisions API does not generate text. It scores up to 1,000 typed questions against one shared input and returns calibrated probabilities. For developers, this targets the small branching calls inside agents: routing, policy checks, tool selection and grading. Google prices it at $0.10 per 1M input tokens, with zero output tokens billed.
TL;DR
- Size: Context is 262,144 tokens (256K) across shared input and all questions.
- Runs on: Google-hosted API only (Gemini Developer API, v1beta). No open weights or self-hosting.
- Performance: Leads 7 of 9 accuracy benchmarks in Google’s table, with the lowest calibration error.
- Best: 92.2% on JevBench (public split).
- Worst: 58.7% Macro-F1 on Amazon ESCI, its lowest score. It also trails Jev 1.14 on When2Call (79.9% vs 82.0%).
- Bottom line: Best: multimodal, calibrated decisions in 150-250 ms. Worst: not the cheapest; Jev 1.14 costs $0.68 vs $0.88 on the shared suite.
What is the Gemini Decisions API?
Gemini Decision 1 is a specialized multimodal model from Google built for closed-set decisions. It reads text, images, video, audio and PDFs in one prefill pass. Output token usage is always 0. Your code branches directly on the returned numbers.
Each request has 3 parts: a model ID, a shared input, and a map of questions. The docs define 3 question types:
noul(yes/no): returns the probability of yes, bounded between 0.01 and 0.99.choice: picks 1 of 2 to 255 options and returns normalized probabilities for every option.score: rates input on 2 to 255 ordered rubric levels and returns a probability-weighted score.
Every answer carries a confidence value from 0 to 1. Google suggests routing low-confidence cases to a fallback or a human reviewer. Questions are scored in parallel and cannot see each other.
How does Gemini Decision 1 work?
Standard models decode an answer token by token, then your code parses it. Gemini Decision 1 instead scores your candidate options in one forward pass. Google reports median model compute of 50-100 ms and 150-250 ms end to end. That holds for inputs up to 16K tokens and stays nearly flat from 1 to 100 questions.
Billing follows the same design. The shared input is billed once per request, however many questions you ask. Each question adds only its own instruction and criteria tokens. There are no thinking-token, transcription or OCR charges.
‘}).join(”);setTimeout(function(){bars.querySelectorAll(‘.fil’).forEach(function(f){f.style.width=f.dataset.w+’%’})},30)}
function mk(el,vals,fmt,mx){el.innerHTML=M.map(function(m,i){return ‘
‘}).join(”);setTimeout(function(){el.querySelectorAll(‘.fil’).forEach(function(f){f.style.width=f.dataset.w+’%’})},30)}
function cost(){mk(R.querySelector(‘#gda-ece’),[0.0250,0.0494,0.0721,0.2104],function(x){return x.toFixed(4)},0.2104);mk(R.querySelector(‘#gda-cost’),[0.88,0.68,0.93,4.14],function(x){return ‘$’+x.toFixed(2)},4.14)}
var rg=R.querySelector(‘#gda-r’);rg.oninput=function(){var t=+rg.value;R.querySelector(‘#gda-tok’).textContent=t.toLocaleString(‘en-US’);R.querySelector(‘#gda-out’).textContent=”$”+(t*1e6/1e6*0.10).toLocaleString(‘en-US’,{maximumFractionDigits:0})};
R.querySelectorAll(‘.lm’).forEach(function(l){l.onclick=function(){l.classList.toggle(‘f’)}});
bench(cur);window.addEventListener(‘load’,fit);window.addEventListener(‘resize’,fit);setTimeout(fit,300);
})();