TAG40@@
:::message
🐹🦜 The two featured in this article

• 🐹 Mocchi (hamster)… AI is still studying. The student role, naively asking questions such as “What does that mean?”
• 🦜 Kinako (budgerigar)… The explanatory role, using AI to handle research. The teacher role, gently teaching through deep dives

This article is written as a transcript of a conversation between the two. The emoji + name at the beginning of each statement identifies the speaker.
:::

🚀 What Is the Gemini Decisions API

🐹Mocchi: I heard Google released another new API? What does it mean to have an API that doesn’t write text?

🦜Kinako: That’s the interesting part. Rather than generating the answer as text, the Gemini Decisions API scores candidate options and returns probabilities and confidence scores. The model is the new Gemini Decision 1 (gemini-decision-1), and it can be used from Google AI Studio via the Gemini API.

🐹Mocchi: What do you use that for?

🦜Kinako: It’s for “decisions that involve choosing from a closed set of options,” such as tool selection, policy checks, model routing, and context scoring. You can submit multiple questions in a single request, and there are three question types.

Key Points

• There are three question types : yes/no (the probability that the condition holds), choice (select from 2–255 options), and score (score using a tiered rubric). Every answer comes with a confidence score from 0 to 1, so only ambiguous cases can be routed to a human or a fallback
• Input and scale: Text, images, audio, video, and PDFs can be passed together as a single input (up to 256K tokens). A single request can evaluate up to 1,000 questions against the same input
• Speed and pricing: End-to-end latency is 150–250 ms. The price is $0.10 per 1 million input tokens, with $0 charged for output tokens. AI Studio also has a free tier
• Official comparison: According to the official blog, it is approximately 5 times cheaper than Gemini 3.5 Flash-Lite + Interactions API in shared benchmarks. Since the benchmark figures are based on Google’s own published results, it would be better to validate performance in your actual use case before deciding whether to adopt it
• Use cases: It is used for model-tier routing of queries in Gemini Notebook, reducing latency for routine inquiries by up to 40%. It is also reportedly used as a security layer for MCP tools in Google Search

🐹Mochie: I see, so it’s specialized for “choosing” an answer rather than “writing” one

🦜Kinako: That’s right. There are cases where it is faster and cheaper than calling a generative model every time, so anyone building a routing layer should definitely check it out

📌 Official primary source

What is the Gemini Decisions API? The official announcement explained in a 3-minute news flash

Please check the linked page for details of the official announcement

🦜 Kinako and Mochie’s Deep Tech Dive explores the latest technology centered on AI and LLMs in depth through conversations between Mochie the hamster and Kinako the budgerigar. Subscribe to the channel and stay tuned!
▶ YouTube channel here