Confirmed 11 Nov 2026 · Labs A one-minute read

Google shipped a model that only decides, priced at $0.10 per million input tokens.

On November 11 Google launched the Gemini Decisions API and Gemini Decision 1. Caveat: Google published no accuracy figures, only the claim that it leads its own benchmarks.

Until now

Agents decide by generating text and then parsing it.

Now

Deciding is a separate priced product, billed only on input.

Why it matters

Every agent makes hundreds of small choices: which tool, is this allowed, which model. Google has made that layer a separate product, charges only for input, and reports 150 to 250 milliseconds end to end. Cheap, fast judgment makes more agent steps affordable, and more steps means less human review.

Established

Gemini Decision 1 is live in the Gemini API and AI Studio.
One request takes up to 1,000 questions over 256,000 tokens.
Pricing is $0.10 per million input tokens, with output free.

Not established

That it is more accurate. No accuracy numbers were published.
That the 40% latency gain holds outside Google's own notebook.
That confidence scores are reliable. Calibration error is claimed, not shown.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Google published "Gemini Decisions API: Fast, calibrated multimodal decisions" on blog.google on November 11, 2026 at 18:00 Universal Time, written by Min Choi, a product manager at Google DeepMind, and Felix Yu, a principal scientist there. We read it there. The endpoint scores a fixed set of candidate answers against one shared input instead of generating text: "Gemini Decision 1 scores your candidate options in a single forward pass." The model is named gemini-decision-1 and is reachable through the Gemini API and Google AI Studio, with Python, JavaScript and REST guides and the google-genai software kit at version 3.2 or later. Three question types are offered: a yes-or-no form returning the probability a condition holds; a multiple-choice form taking 2 to 255 options and returning normalised probabilities; and an ordered rubric form starting at 0 that returns a weighted score with per-level probabilities. Every answer carries a confidence score from 0 to 1, which Google frames as a route to a fallback or a human reviewer. Inputs may be text, images, audio, video and PDFs, up to 256,000 tokens covering the shared input and all questions together, and one request may carry up to 1,000 questions. Google reports 150 to 250 milliseconds end to end, of which 50 to 100 milliseconds is model compute, and says latency stays nearly flat from 1 to 100 questions for contexts up to 16,000 tokens. Pricing is $0.10 per 1 million input tokens with no charge for output tokens, no variable thinking-token charge, and no separate audio transcription or character recognition charge; the shared input is billed once per request, and each question adds only its own instruction tokens. A free tier is offered in AI Studio. Google says the model "leads across the suite, with the lowest calibration error" on agent routing, reasoning, ranking and multimodal benchmarks, and that the product is nearly 5 times cheaper across that suite than Gemini 3.5 Flash-Lite on the Interactions API. Named internal uses: model routing in Gemini Notebook, reported at up to 40% lower latency for everyday queries; wireframing in Stitch; emoji selection from more than 200 options in CC from Google Labs; and a safety layer in Google Search that screens tool calls for adversarial behaviour. Not published: any accuracy or calibration figure, the benchmark names, or the model's size. Listed as unavailable: context caching, batch inference, priority inference, conversation history and tool-call blocks.

How it got here

Oct 8
Google Cloud gives its enterprise agents their own Workspace accounts and directory entries, making an agent an addressable member of a company. Edition 047.
Nov 5
OpenAI releases GPT-6 Cyber to defenders first. Capability is increasingly shipped as a narrow product rather than one general model. Edition 076.
Nov 9
Google launches 2 AI chips, one built to protect model weights while in use, pushing more of the stack into purpose-built parts. Edition 083.
Nov 10
OpenAI puts business agents into ChatGPT and Mercor prices the human work behind evaluation at $20 billion. The judging layer becomes a market. Editions 086 and 087.
Nov 11
Google ships the Gemini Decisions API and gemini-decision-1 at $0.10 per 1 million input tokens, output free, up to 1,000 questions per request.

Who has to react

Every agent framework that pays a general model to make small choices; OpenAI and Anthropic, whose cheapest models are now the comparison; OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, named as integrations; evaluation vendors that sell judging; and the teams who set how often a human sees an agent's decision.

What to watch

Would strengthen itOpenAI or Anthropic ships a decision or classification endpoint that bills input only, or Google reports paid volume for the Decisions API, by June 30, 2027.
Would weaken itGoogle retires gemini-decision-1 or folds it into a general model, or publishes no accuracy figures and names no customer outside Google, by December 31, 2027.
Would strengthen itAn independent evaluation shows gemini-decision-1 beating a general model of similar cost on routing or policy checks, by September 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityEasy

PersistenceYear

Cost is medium. Training a separate decision model and standing up a new endpoint is real engineering, but it is one API among many and no capital is committed. Commitment is medium: Google has published prices and documentation, which are hard to withdraw quietly, yet the model sits in the developer API alone, with Gemini in Google Cloud listed as coming later. Reversibility is easy. Google has retired preview endpoints before and nothing here is contractual. Persistence is a year rather than structural, because the pricing shape, zero output cost, is the part competitors can copy within weeks, and probably will.

The words

Inference: The step where a trained model produces an answer. It is billed per token and is the running cost of any AI product.
Calibration error: The gap between a model's stated confidence and how often it is right. Low error means a 70% score really means about 70%.
Model routing: Sending each request to the cheapest model that can handle it. The router must itself be fast and cheap, or it eats the saving.
Forward pass: One run of the model over its input. Scoring options in a single pass avoids generating text and then parsing it.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.