Google has launched the Gemini Decisions API, an endpoint built on a new model, Gemini Decision 1, that answers closed questions instead of writing text. It is priced at $0.10 per million input tokens, with no charge for output tokens. OpenAI and Salesforce each shipped a decision model in the past two months; Google’s announcement puts a per-token price next to its latency figures, which Salesforce’s announcement of its own model did not.
What Google announced
In a November 11 post, Google said the Decisions API is a new endpoint in the Gemini API, available in Google AI Studio under the model name gemini-decision-1. A request carries one shared input of up to 256,000 tokens, which can mix text, images, audio, video and PDFs, plus up to 1,000 independent questions about that input.
The model handles three kinds of question. It estimates the probability that a yes or no condition holds. It picks one answer from a list of 2 to 255 choices and returns a probability for every candidate. It grades an input against an ordered rubric and returns a weighted score with a probability for each level. Every answer comes with a confidence score from 0 to 1, which Google says lets an application send unclear cases to a fallback model or a human reviewer.
Google says Gemini Decision 1 scores the candidate options in a single forward pass instead of generating an answer token by token. It gives a median of 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end, for inputs up to 16,000 tokens, and says that latency stays nearly flat from 1 to 100 questions per request.
How the bill works
The pricing is the part a finance team will read first. Google charges $0.10 per million input tokens and $0 for output. The shared input is billed once per request however many questions are asked; each question adds only the tokens of its own instruction and criteria. Google says there are no thinking-token charges and no separate fees for audio transcription or OCR. On Google’s own benchmark suite, Decision 1 through the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API.
A no-cost tier is available in Google AI Studio for prototyping. Google says integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, and that support for Gemini in Google Cloud is coming soon. That last point matters for enterprise buyers: today the endpoint sits in the developer-facing AI Studio, and the Google Cloud route that most procurement and data-residency reviews depend on has no date yet.
Where Google already uses it
Google lists four internal uses. Gemini Notebook uses the API as a router that sends each query to a model tier and thinking level, which Google says serves everyday queries with up to 40% lower latency. Google Search uses it as a safety layer that checks MCP tool traffic for adversarial behavior. Stitch and CC, two Google Labs products, use it for interface edits and emoji selection.
Three vendors, three ways of describing the same layer
Decision models arrived fast. TypeSafe AI launched Jev in mid-September, and TypeSafe then raised $870 million at a $7.5 billion valuation. Salesforce’s November 6 post on its own model says OpenAI made its Decisions API generally available about a month after Jev appeared.
Salesforce’s version, Koa Decisions, routes Agentforce requests in about 26 milliseconds, against about 1.4 seconds for a general-purpose LLM, according to Salesforce’s own measurement. Salesforce built it on Google’s open-weight Gemma model, added image and audio support, and says it plans to extend it into flows, Apex and prompt templates, plus an API for developers. Its post gives no price for decision-model calls.
Put the two posts side by side and the numbers do not line up neatly. Salesforce reports a routing latency inside its own platform. Google reports model compute and end-to-end latency across multimodal inputs. Neither post states accuracy on the same task. A buyer comparing them is comparing two vendors’ self-reported results on different tests, and should treat the figures that way.
What Google adds is a published unit price. Salesforce sells its decision model as part of Agentforce, where the cost sits inside the platform contract. Google sells Decision 1 by the token, which means a team can work out the cost of a routing or policy check before it signs anything.
What it means for the buying committee
The economic case for decision models is a change in how an agent’s work is billed. Much of what an agent does is choose: which tool, which team, which model tier, whether a request breaks a policy. Running each of those choices through a general-purpose LLM means paying for generated tokens and waiting for them. A decision model bills the input, returns a short structured answer and a confidence score, and leaves the expensive model for the steps that need it.
That fits the cost pressure AITech has tracked on frontier models. When Google launched Gemini 4 Argon, it did so with introductory pricing that ends November 30, and Ramp’s AI Index showed blended token prices falling while volume rose. A routing layer that sends fewer requests to the large model is one of the few levers a platform team controls directly.
The confidence score also has a governance use. If every routing or policy decision carries a number, a team can set a threshold below which a person reviews the case, and log the score as evidence of why an agent took a step.
What to ask before adopting it
Start with deployment. Ask Google when Decision 1 will be available through Gemini in Google Cloud, and under which data processing terms, since AI Studio terms may not match what the company has signed for Vertex or Workspace.
Next, test accuracy on your own decisions. Google’s benchmark claims and Salesforce’s latency figure come from the vendors. Run a sample of real routing or policy questions through the decision model and through the LLM it would replace, and compare both cost and error rate.
Then set the confidence threshold on purpose. Decide which decisions are allowed to proceed automatically at what score, who reviews the rest, and where the score is logged.
Finally, check lock-in. Salesforce’s model runs inside Agentforce; Google’s runs through the Gemini API and the third-party frameworks it lists. If routing logic is written against one vendor’s question format, moving it later is a rewrite, so keep the decision definitions in your own code.
Source: Google, Introducing the Gemini Decisions API. Additional source: Salesforce, Meet Koa Decisions.
