3BQRIIINO. INDEPENDENT TECH JOURNAL
← All stories

Google / Gemini / AI Models / AI News · November 12, 2026

Google's Gemini Decision 1 Is a Model That Never Writes a Word — It Just Decides

Google launched the Gemini Decisions API on 11 November 2026, powered by the new Gemini Decision 1 model: up to 1,000 yes/no, multiple-choice or rubric questions answered in a single pass, with calibrated confidence scores, in 150–250 ms, at $0.10 per million input tokens and no output charge.

Google's Gemini Decision 1 Is a Model That Never Writes a Word — It Just Decides

Google has released a model that never writes a word. Gemini Decision 1, announced on 11 November 2026 as the engine behind the new Gemini Decisions API, reads text, images, audio, video or PDFs and answers up to 1,000 yes/no, multiple-choice or rubric questions about them in a single pass — returning probabilities and a confidence score for each, in about a quarter of a second. It is the third decision-model launch in under two months, and the first from Google.

The idea is simple enough that it sounds obvious once stated: most of what an AI agent does all day is not writing. It is choosing. Which tool should run next? Does this content break the policy? Which model should handle this query? How good is this draft, on a scale? A general-purpose model answers those questions by generating sentences, which then have to be parsed back into a value the code can act on — slowly, expensively, and with a confidence nobody can measure. A decision model skips the prose entirely.

How it works

Instead of generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass. You send one shared multimodal input — up to 256K tokens of text, images, audio, video or PDF — plus a list of up to 1,000 independent questions, and the model returns a probability for every option, with a confidence score from 0 to 1. The input is billed once per request no matter how many questions you ask; each question adds only its own instruction and criteria tokens.

There are three question types, and the names tell you the whole design:

The confidence score is the feature that matters most in practice. Every answer carries one, so code can route the clear cases automatically and send only the ambiguous ones to a fallback or a human reviewer. That is the difference between a classifier you bolt on and a control plane you can build on.

The numbers — all Google's own

Google's announcement quotes 150–250 ms end-to-end latency (50–100 ms of model compute), nearly flat from 1 to 100 questions per request, across multimodal inputs and contexts up to 16K tokens. Pricing is $0.10 per million input tokens, with zero output tokens billed — there are no output tokens, no thinking-token bills, and no separate audio transcription or OCR charges. Google also claims the Decisions API is nearly 5× cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API, and that Decision 1 leads its benchmark suite — spanning agent routing, reasoning, ranking and multimodal understanding — with the lowest calibration error.

Treat those figures the way you would any launch-day claim: they are Google's measurements on Google's benchmarks, with no independent verification published yet. The price, at least, is a fact you can check on the rate card — and it is identical to the $0.10 per million input tokens OpenAI lists for its own Decisions API. The latency and benchmark leadership are claims until someone else measures them.

A race nobody announced

Decision models arrived as a category almost all at once. TypeSafe launched Jev on 15 September, OpenAI made its Decisions API generally available on 22 October, and Microsoft followed with Microsoft-Decision-1, built on Alibaba's Qwen. Google's entry, a day after it opened Gemini 4 Carbon to paying users, completes the set: every major lab now has a model whose entire job is to pick, not to write.

That timing is not a coincidence. Agents have become the product, and agents spend most of their time deciding — which tool, which model, which branch of the workflow. The bottleneck in an agent loop is no longer the quality of one answer; it is the cost and latency of hundreds of small judgements per task. A model that answers a thousand questions in one pass for ten cents per million tokens is infrastructure, not a demo.

Google is already using it on itself

The most persuasive part of the announcement is not the benchmark chart. It is the list of Google products already running on the API:

That is the pattern to watch: decision models first show up inside the lab's own products, as plumbing, and only later as something developers are asked to build with.

What to know before building on it

Gemini Decision 1 is available now through the Gemini API in Google AI Studio, with a no-cost tier for prototyping. It is a v1beta API today, and Google says support in Gemini on Google Cloud is coming soon. Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs.

Two practical notes. First, the 1,000-question figure is a ceiling, not a recommendation: Google's own latency numbers are quoted for contexts up to 16K tokens, and the flat-latency claim covers 1 to 100 questions per request. Second, the free tier exists for prototyping, and Google's pricing page notes that free-tier data can be used to improve its products while paid use is not — the usual trade-off, worth knowing before you point it at anything sensitive.

Why this launch matters

The frontier race in November 2026 has two tracks, and they are easy to confuse. One is the race for the biggest, smartest general model — the track Gemini 4 Carbon and Claude Fable 5.2 run on. The other is the race for the small, fast, cheap models that sit underneath agents and make the thousands of small decisions a single task requires. Decision 1 is the second track, and it may matter more for what gets built next: the cost of deciding is what determines how many agents a product can afford to run.

Google did not invent the category. It joined it, a month after OpenAI and two months after TypeSafe — and it joined with the largest distribution of any of the entrants, through AI Studio and, soon, Google Cloud. The question now is not whether decision models are real. It is whether developers start designing agents around them, or keep paying a flagship model to write paragraphs they immediately throw away.

Sources and classification

Classification: Official Announcement (Google, 11 November 2026). Launch facts, availability and pricing are from Google's own announcement; benchmark, latency and cost-comparison figures are Google's own claims, attributed as such above, with no independent verification published yet. Checked 12 November 2026.

Prefer reading in Arabic? You can translate this article into Arabic with Google Translate.