Google's Gemini Decisions API uses a model that writes zero text, scoring up to 1,000 yes/no, choice, and rubric questions per call at $0.10/M tokens.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Google shipped a model this week that refuses to write a single sentence. The Gemini Decisions API, powered by a new model called Gemini Decision 1, doesn't generate text at all. You send it a pile of context and a checklist of yes/no, multiple-choice, or rubric questions, and it answers all of them in one pass with calibrated probabilities. Zero output tokens, which means zero output cost.
This isn't a side feature bolted onto Gemini. It's a dedicated endpoint, a dedicated model, and a direct shot at a category that didn't exist two months ago.
It's a single REST endpoint — POST https://generativelanguage.googleapis.com/v1beta/decisions — that takes shared input (text, images, audio, video, or PDFs, up to 256K tokens) plus a map of up to 1,000 questions, and scores every question against that input in one forward pass. According to Google's own docs, questions can't see each other, but asking more of them barely adds latency, since the model isn't decoding tokens for each one.
There are three question types, and the naming is a little goofy:
true_description and false_description fields for when "yes" is ambiguousEvery answer ships with a confidence score from 0 to 1, which is the part that actually matters. A routing system doesn't need a model that's always right, it needs a model that knows when it isn't, so the uncertain 10% can get escalated to something smarter.
Normal Gemini calls generate a response token by token, then you parse it, hope the JSON is valid, and pay for every token it wrote on the way to "yes." Decision 1 skips the writing entirely. It's scoring predefined options against the input in a single pass, the same way a classifier head works, except it's built on a frontier multimodal model so it actually understands the image or PDF you handed it.
That's the whole pitch: structured output without the parsing tax, at a fraction of the latency and cost of a full chat completion. Google says it runs 150-250ms end to end, with 50-100ms of that being actual model compute. At $0.10 per million input tokens and nothing for output, Google claims that's roughly 5x cheaper than running the same check through Gemini 3.5 Flash-Lite.
The three buckets Google keeps pointing to:
Routing. Classify a request, a ticket, or a user message, then send it to the right queue, department, or model tier. Google says this is already live inside Gemini Notebook for picking which model handles a given query, and it's cut latency there by up to 40%.
Tool selection. Instead of letting an agent's main model burn tokens deciding which of 12 tools to call, a cheap decisions pass can narrow it down first, or flag whether the situation even needs escalation at all.
Grading. Score an agent trace, a support reply, or a piece of content against a rubric, fast enough to run on every single item instead of a sample.
There's a batch angle too: up to 1,000 independent questions against one shared input, useful for things like scanning a stack of delivery photos for damage, or running a full compliance checklist against one document in a single call instead of a thousand round trips.
This is the part that makes the launch interesting instead of just another API. Google isn't first here, not even close. TypeSafe AI's Jev kicked off this whole category in mid-September, hit a million users in days, and raised $870 million at a $7.5 billion valuation three weeks later. OpenAI followed with its own Decisions API, built on GPT-6 Luna, going from DevDay preview on September 29 to general availability on October 22, also at $0.10 per million tokens.
So by the time Gemini Decision 1 showed up on November 11, it was the third entrant into a market that's barely eight weeks old. Google's own benchmark numbers admit this isn't a clean sweep: Decision 1 leads on 7 of 9 accuracy benchmarks against Jev 1.14 and GPT-6 Luna, and roughly halves Jev's calibration error (0.0250 vs 0.0494 Top-1 ECE). But Jev still wins on Mind2Web Multimodal (86.8% vs 86.2%) and, notably, on When2Call — the benchmark that's literally about knowing when to invoke a tool, where Jev scores 82.0% against Decision 1's 79.9%. If tool selection is your main use case, that's the one number worth sitting with before you pick a model on reputation alone.
And on price, Google isn't even competitive. Jev lists at $0.042 per million input tokens, less than half of Decision 1's $0.10, and on the seven benchmarks both models share, running the equivalent workload through Jev costs $0.68-$0.88 against Decision 1's bill. Where Google actually pulls ahead is multimodal: Jev and GPT-6 Luna don't score on video benchmarks at all, while Decision 1 hits 69.1% on Video-MME without subtitles. If your decisions involve audio, video, or PDFs, Google's the only one of the three that's even in the conversation.
I think this category is real, and I said as much when I wrote about the Jev-plus-Opus routing pattern a few weeks back. Running every judgment call through an expensive reasoning model was always a waste for the 80-90% of cases that are genuinely easy. What Google's launch confirms is that this isn't a Jev gimmick, it's a structural gap in how LLM APIs charge you: full chat completions bill for prose you never read, on every single classification call, forever.
That said, I wouldn't switch off Jev just because Google showed up with a bigger brand name. The benchmarks don't say Decision 1 is better, they say it's better at some things and worse at others, and $0.10 versus $0.042 is a real gap once you're running millions of these a day. If you're already deep in the Gemini ecosystem and need video or PDF support, Decision 1 is the obvious pick. If you're purely doing text classification and tool routing at scale, Jev's price and its When2Call lead make it hard to justify paying more than double for a worse fit. And if you're on OpenAI already, their Decisions API being a month further along in GA than Google's preview-stage launch is worth factoring in too.
The bigger lesson, same one from the Jev post, still applies: confidence scores are not accuracy. Thresholds matter more than the model you pick. Whatever you choose here, tune it on your own labeled data before you let it make unsupervised decisions in production.
It's a Google API endpoint powered by a new model, Gemini Decision 1, that answers yes/no, multiple-choice, or rubric-scoring questions about multimodal input (text, images, audio, video, PDFs) without generating text. It returns calibrated probabilities instead of a written answer, and bills zero output tokens since it writes nothing.
$0.10 per million input tokens, with no charge for output tokens since the model generates none. Google says that's roughly 5x cheaper than running the same classification task through Gemini 3.5 Flash-Lite. It's double the price of Jev's $0.042 per million tokens, the main competing decision model.
Google reports Decision 1 beats Jev 1.14 and GPT-6 Luna on 7 of 9 accuracy benchmarks and has about half Jev's calibration error. Jev still wins on Mind2Web Multimodal and When2Call (tool-use timing), and is cheaper on every shared benchmark. Decision 1 is the only one of the three with real video and audio support.
Three: "noul" for yes/no probability estimates (0.01-0.99), "choice" for picking among 2-255 labeled options with a probability per option, and "score" for rating against an ordered rubric of 2-255 levels. You can ask up to 1,000 questions against one shared input in a single request.
Yes. It accepts text, images (PNG, JPEG, WebP, HEIC, and more), audio (WAV, MP3, AAC, and others), video (MP4, MOV, WebM), and PDFs up to 50MB inline, all within a 256K-token context window. It's the only major decision model of the three (Google, Jev, OpenAI) that currently scores on video benchmarks.
Yes, it launched November 11, 2026 in preview through Google AI Studio and the Gemini API (v1beta), with a free tier for testing. It's also live on OpenRouter as google/gemini-decision-1 from day one. Google Cloud/Vertex AI support is coming later.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Jane Street open-sourced a single SKILL.md file that turns Claude Code into a strict Socratic tutor. Here's how the skill map, diagnostic, and final exam work, and how to install it.
Gemini 4 Hydrogen costs $1.50/$7.50 per million tokens, beats Argon's intelligence score, and runs Orchestrator in Antigravity. Here's what's true and what's hype.
Libghostty 1.0 freezes a stable C ABI for Ghostty's terminal engine, already used by JetBrains IDEs (12.5M users), Neovim, KDE, and Turborepo.
Comments