Google’s Decisions API scores 1,000 questions at once for $0.10 per million tokens

Google has launched the Gemini Decisions API, a new endpoint that scores up to 1,000 closed-set questions against one multimodal input in a single pass. It runs on a new model, gemini-decision-1, and bills $0.10 per 1M input tokens with no output charges.

One input, up to 1,000 questions, probabilities back

Gemini Decision 1 does not write an answer token by token. According to Google, it scores the candidate options in a single forward pass and returns probabilities and confidence scores. Google puts end-to-end latency at 150-250 ms, with 50-100 ms of model compute, and says latency stays nearly flat from 1 to 100 questions per request.

The input can combine text, images, audio, video and PDFs, up to 256K tokens, and the question limit is 1,000 per request. Google lists three question types. The noul type estimates the probability that a yes/no condition holds. The choice type picks from 2 to 255 options and returns normalized probabilities for each. The score type grades against an ordered rubric starting at 0, returning a weighted score and per-level probabilities. Every answer carries a confidence value from 0 to 1, intended so code can send ambiguous cases to a fallback or a human reviewer.

The input field uses the same content block format as the Interactions API. Google’s example in the google-genai Python SDK checks a photo of a breakroom counter against a yes/no question, a three-way cleanup choice and a three-level severity rubric in one client.decisions.create call, using the v1beta API version.

Pricing that bills the shared input once

The model costs $0.10 per 1M input tokens and $0 for output. The shared input is billed once per request regardless of the number of questions; each question adds only its own instruction and criteria tokens. Google says there are no variable thinking-token charges and no separate audio transcription or OCR fees. A no-cost API tier in Google AI Studio is available for prototyping, and the product is described in Google’s social posts as a preview.

Google also claims that, across a shared benchmark suite, the API is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API, and that Gemini Decision 1 leads on agent routing, reasoning, ranking and multimodal tests with the lowest calibration error. The post shows benchmark charts but the article text gives no per-test scores, so those claims are Google’s own and cannot be checked from the announcement.

Outside tooling is arriving too: Google names eight third-party platforms, among them OpenRouter, LangChain, Vercel and LiveKit, where support is being added alongside the google-genai SDKs. Support in Gemini in Google Cloud is listed as coming soon, so enterprise customers on Google Cloud cannot use it there yet.

Google’s own products already use it as a router

Google names four internal uses. Gemini Notebook routes queries to a model tier and thinking level, which Google says gives up to 40% lower latency on everyday queries. Stitch uses it for a rapid canvas that builds playable wireframes. CC from Google Labs picks emoji reactions from more than 200 options. Google Search uses it as a safety layer for MCP tools, detecting adversarial behavior, though Google gives no figure for how well it works.

Who benefits: Teams paying for routing and grading calls

The clearest users are developers who today spend generative calls on small decisions: which model to send a prompt to, whether a document meets a policy, whether retrieved context is relevant. Google’s own research suggests why that matters. A test published on November 3 sent every task to Flash-Lite and escalated only those a classifier flagged as multi-step to Gemini 3.8 Flash on low thinking. On its 50-question coding set the router solved all 50, though Google noted that the benchmark code was not public and that difficulty labels were assigned by hand. The Decisions API packages that kind of gate as a hosted endpoint.

Google is not alone in the format. Hugging Face recently released AnyDecisionModel, a Swift package with the same three question shapes – yes/no, choice and rating – that runs on-device and caps next-token backends at 26 options per choice. Google’s hosted version allows 255 options, but sends data to its servers, while the Hugging Face tool can run locally on Apple silicon. Hugging Face also cautions that its confidence values show how concentrated the probabilities are, not whether an answer is correct. The same caveat plausibly applies here: Google calls its scores calibrated, but teams should measure error rates on their own data before letting a confidence threshold decide who gets a human review.

The source leaves several things open. It gives no rate limits, no regional availability, no paid-tier terms beyond the token price, and no scores for the benchmark suite in text. It also lists a 256K-token input limit while the latency claim covers contexts up to 16K tokens, so speed on very long inputs is not established. Google’s launch posts frame the product as a preview, which means pricing and behavior may change.

Developers can start by creating a key in Google AI Studio and following the Python, JavaScript and REST guides in the documentation. Google also lists three demo apps: Doorstep Parcels, Speed Sketch and Find the Zoo Keeper.