Gemini Decisions API
Google has launched its Gemini Decisions API, a new service offering rapid, calibrated multimodal decision-making powered by the Gemini Decision 1 model. This move positions Google as a direct competitor to existing 'decision models' like Jev and OpenAI's offerings, sparking intense discussion among Hacker News users about market dynamics, technical prowess, and Google's product longevity.
The Lowdown
Google introduces the Gemini Decisions API, a new endpoint for fast, closed-set decisions driven by their Gemini Decision 1 model. This API provides purpose-built primitives for bounded choices, such as tool selection, policy checks, model routing, and context grading.
- It processes multimodal inputs including text, images, audio, video, and PDFs (up to 256K tokens) and evaluates up to 1,000 independent questions in a single forward pass.
- Unlike generative models, Gemini Decision 1 scores candidate options directly, returning probabilities and confidence scores with low latency (150-250 ms end-to-end) and a cost-effective pricing model ($0.10 per 1M input tokens, zero output tokens billed).
- It supports three decision types: 'nool' (yes/no), 'choice' (multiple choice), and 'score' (ordered rubric), each providing confidence scores to handle ambiguous cases.
- Teams within Google are already leveraging the API for dynamic model routing in Gemini Notebook, rapid design iteration in Stitch, emoji selection in CC, and as a safety layer for Search tools.
- The API is available through the Gemini API in Google AI Studio, with integrations rolling out across various platforms like OpenRouter, LangChain, and Vercel.
This release marks Google's formal entry into the specialized decision model space, offering a robust tool for developers seeking high-performance, cost-efficient AI-driven decision-making capabilities.
The Gossip
The Jev Effect: Google's Entry into Decision Models
Many commenters noted the striking similarity to Jev's API and discussed Google's rapid entry into this emerging market. There was a sentiment of surprise that it took a startup like Jev to highlight the utility of such a product, prompting the larger players to follow. The conversation included comparisons of pricing and performance claims, with some skepticism about self-reported benchmarks. Concerns were raised about Google's 'we do it too' strategy and its potential impact on smaller innovators like Jev.
Practical Potentials & Purposeful Puzzles
Users explored current and potential applications for decision models. Some shared examples of using such APIs for labeling training data and as a control plane for AI agents. Others detailed their home-rolled systems for file indexing and tagging using similar principles. However, a recurring question was a fundamental query about the overall utility and clear demonstration of where these decision models truly shine.
Probabilistic Ponderings: Calibration and Underpinnings
A significant technical discussion revolved around how these models calculate probabilities and the crucial importance of probability calibration for reliable decision-making. Commenters debated whether standard LLM log probabilities are sufficient or if specialized tuning is required, especially after instruction fine-tuning and RLHF. Some pointed to Natural Language Inference (NLI) as a historical precedent for similar classification tasks with smaller models.
Google's Graveyard: Trust and Longevity Concerns
Consistent with common Hacker News discussions about Google, many expressed skepticism regarding the long-term viability and commitment to new Google products, citing the company's history of 'killing' services. While some feared this new API might meet a similar fate to 'Jules,' others argued that enterprise-focused APIs, especially those generating revenue, typically have a much longer lifespan and different support model compared to consumer-facing or experimental projects.