Google has launched Gemini Decision 1 — a new AI model built not to generate text or images, but to choose. It’s live in Google AI Studio as part of the new Gemini Decisions API, and it answers up to 1,000 yes/no or multiple-choice questions against a single multimodal input in one forward pass.
No streaming. No token-by-token output. Just decisions, with confidence scores from 0 to 1. And it’s priced to undercut Gemini 3.5 Flash-Lite by nearly five times on shared benchmarks, with zero output tokens billed.
That’s the core shift: this isn’t another chatbot layer. It’s infrastructure for routing, grading, and triaging, built into the Gemini stack from day one. The model accepts text, images, audio, video, and PDFs, all at once, natively. A developer could feed it a 10-second call recording plus a customer’s account ID plus a screenshot of their last support ticket, then ask 27 policy-compliance questions across that bundle.
Or pair a damage photo with a scanned insurance claim PDF and run 400 validation checks in one request. Input size caps at 256K tokens. Image and video resolution is adjustable to balance detail against cost. Google says its internal teams are already using it as a low-latency control plane, replacing custom classifiers and generative-model round trips for tasks like agent routing, context grading, and tool selection.
Four internal use cases are documented, though no names or verticals are given. Availability is immediate: developers can create an API key in Google AI Studio and test it via three demo apps, Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. SDK support is live for Python, JavaScript, and REST.
Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. Google Cloud integration is “coming soon”. Pricing details are confirmed only in relative terms: Google states the Decisions API is “nearly five times cheaper” than Gemini 3.5 Flash-Lite on its shared benchmark suite, and that billing excludes output tokens entirely.
No per-request, per-token, or monthly-tier pricing is disclosed for India or any other market. No Indian launch event, local pricing, or regional rollout timeline is mentioned in the announcement. This isn’t a consumer product. There’s no app, no UI, no Pixel or Android integration announced.
It’s a backend API, aimed squarely at engineers building decision-heavy workflows: fraud detection pipelines, compliance bots, multimodal QA systems, or real-time routing logic for contact centres. What changes now? Developers who previously stitched together LLM calls, rule engines, and confidence thresholds, often with latency and cost overhead, now have a single, calibrated, multimodal decision endpoint.
Whether that replaces existing stacks depends on how well it handles edge cases, scale, and domain-specific calibration. Google says it leads on “calibration error” across its internal benchmarks. Independent verification will follow. Next: expect early adopters to publish integrations, especially in fintech, insurance, and enterprise support, within weeks.
Google Cloud support arrives later. And if this proves robust, it may quietly become the invisible logic layer behind future Gemini-powered features in Workspace, Android, or even Search, though none of that is promised, priced, or scheduled.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.