Google unveils Gemini Decisions, 1,000 questions per query

Google launches a decision API that returns probabilities instead of text and accepts multimodal inputs with a large context window. Priced at $0.10 per million input tokens and measured with latency between 150 and 250 ms, it is accessible via AI Studio. However, experts are calling for guidelines on its use in production and for evaluating these systems based on their actual reliability, while OpenAI offers a competing service with a narrower range of questions.

Possible Erroneous Judgments with Manipulated Data

Variations in data or deliberately manipulated inputs can lead these APIs to erroneous judgments, warns Lian Jye Su. A confidence score alone does not establish the reliability of an assessment. According to him, processes must remain highly contextualized and rely on a layer of reliable and secure context. He recommends linking models to historical transaction databases and established business policies to evaluate requests against relevant organizational information rather than solely on immediate input. He adds that Google will need to meet corporate data retention and sovereignty requirements without compromising costs and performance. Tulika Sheel emphasizes the need for clearly defined human accountability as more autonomous agents make decisions and act on behalf of users.

A Probability Decision Model with a Window of Up to 256,000 Tokens

Powered by Gemini Decision 1, the API can handle up to 1,000 questions in a single request and return probabilities on defined options rather than generated text. It allows for controlling conditions, selecting from predetermined choices, or scoring inputs according to an evaluation grid to estimate their compliance with criteria. Accepted inputs can be of various types, such as images, audio files, or videos, and fit within a context window of up to 256,000 tokens. The assigned confidence levels can guide automation or decide when to escalate to human validation. Google reports an end-to-end latency of 150 to 250 milliseconds for contexts up to 16,000 tokens, with relatively stable performance for one to 100 questions.

Pricing at $0.10 per Million Input Tokens and Availability

The service is priced at $0.10 per million input tokens, with no output fees. Google specifies that shared input is charged once per request, with each question adding only its own instruction and criteria tokens. The API is accessible via Google AI Studio. Integrations with development frameworks are being rolled out, and support on Google Cloud is planned for the future, with no announced date.

Internal Uses Already in Place in Notebook and Search

Google is already using this technology in Gemini Notebook to route queries to different AI models based on the complexity of each request. According to the company, this approach has reduced latency for daily queries by up to 40%. The API is also used in Search to detect adversarial behaviors involving Model Context Protocol tools, beyond simple model routing.

Competition from OpenAI and Use Cases Considered by Analysts

Beyond query routing, applications are being considered in policy evaluation, risk assessment, and credit scoring, according to Lian Jye Su, who finds the model suitable for unstructured and dynamic inputs. In comparison to OpenAI, whose Decisions API has been generally available since October and limits a request to 400 questions, Google boasts a capacity of 1,000 questions. Both charge $0.10 per million input tokens with no output fees, with OpenAI additionally offering discounted rates for cached inputs. Google’s offering accepts a broader multimodal palette, including audio and video. For Lian Jye Su, integration into a company's development environment weighs more than a surface-level comparison, and developers should assess reliability in real-world conditions. He sees the main benefit in reducing judgment errors and improving decision quality.