What Google actually announced
According to Google's announcement, the Gemini Decisions API is a new endpoint for fast, closed-set decisions, powered by a new model called gemini-decision-1 and available through the Gemini API in Google AI Studio.
The design is deliberate and narrow. Instead of generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass. You pass one multimodal input (text, images, audio, video and PDFs, up to 256K tokens) together with up to 1,000 independent questions. The API returns probabilities and confidence scores in 150–250 ms end-to-end latency, at $0.10 per 1M input tokens with zero output tokens billed.
Three question types are supported: noul (yes/no, with optional true and false criteria), choice (multiple choice from 2 to 255 options, returning the winner and normalised probabilities for every candidate), and score (an ordered rubric, returning a continuous weighted score and per-level probabilities). Every answer carries a confidence score from 0 to 1, so your code can route ambiguous cases to a fallback or a human reviewer.
How it differs from what came before
The interesting shift is architectural rather than cosmetic. Generative APIs ask a model to produce text and let you infer a decision from that text: parse the wording, check the shape, trust the tone. The Decisions API asks the model to produce the decision directly, with a probability attached.
That removes two familiar failure points in legal and marketing workflows: prompt engineering for output format, and post-hoc parsing of free text into a usable field. The source also notes that latency stays nearly flat from 1 to 100 questions per request, which matters when a single document needs forty checks rather than one.
| Dimension | Gemini Decision 1 + Decisions API | Gemini 3.5 Flash-Lite + Interactions API |
|---|---|---|
| Output | Probabilities and confidence scores only | Generated answer text |
| Output-token billing | $0 | Billed |
| Input pricing | $0.10 per 1M input tokens | Not stated in source |
| Latency | 150–250 ms end-to-end; 50–100 ms median model compute | Not stated in source |
| Questions per request | Up to 1,000, nearly flat from 1 to 100 | Not stated in source |
| Cost across shared benchmark suite | Baseline | Nearly 5x more expensive, per Google |
According to Google, Gemini Decision 1 leads across benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. Calibration is the quiet headline: a confidence score that is actually reliable is more useful than a confident paragraph that is not. Compared with building custom classifiers, the source frames the product as replacing training work with option definitions.
Three concrete uses worth testing
1. Intake triage in a small practice. A firm receives enquiries by email, voice note and photographed letters. A single request could attach the message and its attachment, then ask a yes/no question (does this appear to be a new client enquiry?), a choice question (practice area from a bounded list) and a rubric score (urgency). Low-confidence answers route to a human. This is a realistic first automation for a team exploring AI for lawyers without rebuilding its practice management system.
2. Document and claim checking with evidence. The source's own example pairs a scanned PDF claim with damage photos in one shared input. The same pattern applies to conveyancing packs, insurance submissions or compliance files: one input, several checks, each returning a probability. Because the shared input is billed once per request regardless of question count, the economics favour asking many narrow questions rather than one broad one.
3. Creative and marketing asset gating. For founders and creators, a rubric score across a batch of drafts can flag which assets need a human read. Google's own Stitch and CC examples show the pattern for design and emoji selection. Anyone building an agent layer on top of this should read AI agents for lawyers before wiring decisions into client-facing steps.
Who and what this affects
According to the announcement, teams across Google are already using the API as a low-latency control plane: Gemini Notebook as a real-time router, Stitch for rapid design iteration, CC for emoji selection, and Google Search as a safety layer for MCP tools detecting adversarial behaviour and content. Those examples suggest the primary audience is developers building agent systems, not end users.
For legal and professional services, the practical impact is on the glue work: routing, gating, grading and escalation. Work previously done by bespoke classifiers or brittle prompt chains becomes an option definition plus a threshold. That lowers the cost of entry for small firms and solo practitioners, but it also shifts the scarce skill from prompt writing to question design and threshold setting.
There is a commercial angle too. Google states that there are no output-token charges, variable thinking-token bills, or separate audio transcription and OCR costs. For firms currently paying for separate transcription and document processing, that consolidation is at least worth a costing exercise.
What to watch: confidentiality, accuracy, IP and professional limits
Confidentiality. The API accepts audio, video and PDFs, which is precisely the material clients expect to be protected. Before anything reaches a prototype, confirm which tier you are using and what your data terms say. The no-cost tier in Google AI Studio is described by the source as being for prototyping, and that framing should be taken literally.
Accuracy and calibration. A confidence score is not a correctness guarantee. Google claims the lowest calibration error across its benchmark suite, but calibration on benchmarks does not automatically transfer to your documents, your jurisdiction or your language mix. Build a small labelled test set from your own matter files before trusting any threshold.
Intellectual property. Question sets, criteria wording and rubric levels are authored content with commercial value. Treat them as assets, and consider who owns them under your engagement terms; the discipline described in protecting creative ideas before launch applies to internal tooling as much as to client work.
Limits of professional use. The API answers bounded questions; it does not give legal advice or exercise judgement. Rubric scores and probabilities are inputs to a human decision, not substitutes for one. Any client-facing use should keep a named reviewer in the loop, and any marketing claim about automated accuracy should be checked against professional conduct expectations.
How and when to try it
You can start today, according to the source: obtain an API key in Google AI Studio and follow the Python, JavaScript and REST guides in the Decisions API documentation. Google also points to three demo apps, Doorstep Parcels, Speed Sketch and Find the Zoo Keeper, as starting points.
Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode alongside the google-genai SDKs, with support coming soon to Gemini in Google Cloud. A sensible first experiment: take twenty real, anonymised documents, write five bounded questions per document, and compare the returned probabilities against your own answers. That single exercise will tell you more about fit than any benchmark table.
Sources
Frequently asked questions
What is the Gemini Decisions API in plain terms?
It is a Google endpoint that answers bounded questions rather than writing text. You send one multimodal input plus up to 1,000 yes/no, multiple-choice or rubric questions, and it returns probabilities and confidence scores in a single pass. According to Google, latency is 150–250 ms end-to-end and output tokens are not billed.
Can a law firm use it to make legal decisions automatically?
It is better understood as a triage and routing tool. The API returns scores and probabilities for questions you define; it does not exercise legal judgement. A defensible pattern is to automate intake, classification and completeness checks, then route ambiguous or high-stakes results to a qualified reviewer using the confidence score as the trigger.
How does the pricing compare with ordinary generative calls?
According to Google, Gemini Decision 1 costs $0.10 per 1M input tokens with zero output tokens billed, and the shared input is billed once per request regardless of how many questions you ask. Google states this makes it nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
Want to apply this to your own work or your team? Tell me where you stand, and I will reply personally.
Get in touch
