Google launches Gemini Decisions API for fast multimodal decisions, now on Vercel

What Google launched

Google has introduced the Gemini Decisions API, a new endpoint powered by a dedicated model, Gemini Decision 1 (gemini-decision-1), available through the Gemini API in Google AI Studio. According to a Google AI blog post, the API accepts one multimodal input — text, images, audio, video or PDFs, up to 256K tokens — and evaluates up to 1,000 independent questions against it in a single request. Instead of decoding an answer token by token, the model scores the candidate options in one pass through the network, returning probabilities and confidence scores with 150–250 ms end-to-end latency.

Google prices the model at $0.10 per 1M input tokens, with no charge for output tokens. The shared input is billed once per request regardless of how many questions are asked; each question only adds its own instruction and criteria tokens.

Three kinds of questions

The API supports three decision formats. A Boolean type estimates the probability that a yes/no condition holds, with optional descriptions of what true and false mean. A choice type picks from 2 to 255 options, returning the winner plus normalized probabilities for every candidate. A score type grades the input against an ordered rubric, producing a continuous weighted score and per-level probabilities. Every answer includes a confidence value between 0 and 1, which lets application code send ambiguous cases to a fallback path or a human reviewer.

Inputs use the same Content block format as the Interactions API, and modalities can be mixed — for example pairing a call recording with account metadata, or a scanned PDF claim with damage photos. Resolution on image and video inputs can be tuned to trade visual detail against token cost.

Performance and cost claims

Google says Gemini Decision 1 tops its benchmark suite covering agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error of the models compared. Reported median model compute is 50–100 ms, and latency stays nearly flat between 1 and 100 questions per request at contexts up to 16K tokens. Google also claims the model is close to five times cheaper than Gemini 3.5 Flash-Lite on the Interactions API across the shared benchmark suite, helped by zero output-token billing and the absence of separate audio transcription or OCR charges. A no-cost tier is available in AI Studio for prototyping.

Google lists several internal adopters. Gemini Notebook uses the API to route incoming queries to the right model tier and thinking level, cutting latency by up to 40% on everyday queries. The Stitch design tool uses it to generate playable wireframes and apply voice or tap edits almost instantly, CC from Google Labs uses it to pick emoji reactions from a palette of more than 200 options, and Google Search applies it as a safety layer that detects adversarial behavior around MCP tools.

Availability through Vercel

According to Vercel's changelog, the model is now live on the AI Gateway as google/gemini-decision-1 and can be called through the AI SDK, the TypeSafe-compatible API, or the Decisions API. PDF, audio and video inputs work through AI SDK 7, while the TypeSafe-compatible and Decisions APIs currently handle text and image inputs only. Decision calls appear in gateway logs and count toward budgets like any other model request.

Google adds that ecosystem integrations are rolling out across OpenRouter, LangChain, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs, with support for Gemini in Google Cloud described as coming soon.

Why it matters

A large share of agent work — routing requests, checking policies, filtering tools, ranking results — does not need generated prose; it needs a fast, well-calibrated pick among known options. Developers have typically either trained bespoke classifiers or paid full generative latency and token costs to extract such decisions from a chat model. A decision-native endpoint that bills nothing for output and answers in under 250 ms turns those tasks into cheap control-layer calls, and built-in confidence scores make escalation to a fallback or a human reviewer straightforward. With the model already exposed through gateways like Vercel's, it can slot into existing AI SDK stacks with little rework — and if the benchmark claims hold up under outside testing, dedicated decision models look set to become a standard layer between applications and generative models.