Google launched the Gemini Decisions API on November 11, 2026. It is a developer endpoint that does not write text. You send one input and up to 1,000 questions, and it returns a probability and a confidence score for each answer. The model behind it is Gemini Decision 1 (gemini-decision-1). It costs $0.10 per million input tokens and nothing for output. Here is what it does, what it costs, and what the early numbers do and do not show.
What is the Gemini Decisions API?
According to Google’s launch post, the API is built for bounded choices: tool selection, policy checks, model routing, query routing and content grading. Instead of generating an answer token by token, the model scores the candidate options in a single forward pass. A request can mix text, images, audio, video and PDFs as shared input, and Google lists three question types:
- Yes/no: the probability that a condition holds.
- Choice: pick from 2 to 255 options, with a normalized probability for each.
- Score: grade the input on an ordered rubric, with a weighted score and per-level probabilities.
Every answer carries a confidence score from 0 to 1, so unclear cases can be routed to a fallback or a person. Google says it is available through the Gemini API with a playground and API keys in Google AI Studio.
Gemini Decisions API pricing
Pricing as stated by Google and repeated by Implicator and Kingy:
- $0.10 per million input tokens, the same for every input type.
- $0 for output tokens, with no thinking-token charge and no separate audio transcription or OCR fee.
- The shared input is billed once per request. Each question adds only its own instruction and answer options.
- A no-cost tier in Google AI Studio for prototyping. Kingy notes that free-tier data is used to improve Google’s products.
That shared-input billing is the cost lever. Implicator’s worked example, based on Google’s rate, puts a 5,000-token request (a 4,000-token record plus 20 checks) at about $0.50 per 1,000 records, against about $8.10 per 1,000 records if the same 20 checks are sent as separate calls. Kingy’s own arithmetic for a message plus 10 questions (about 2,500 tokens) is roughly $0.25 per 1,000 requests. Those are the outlets’ calculations, not Google figures.
One limit on savings: there is no context caching and no Batch API, so repeated input is billed in full each time, per Kingy and eesel.
Limits and rate limits
- Up to 262,144 input tokens per request (Google rounds this to 256K).
- 1 to 1,000 questions per request.
- Latency of 150 to 250 ms end to end, with median model compute of 50 to 100 ms, and almost no change from 1 to 100 questions for contexts up to 16K tokens. Google advises teams that use the full context or all 1,000 questions to measure their own latency.
- Kingy lists rate limits by tier: the free tier allows 200 requests a minute and 50,000 tokens a minute, so it cannot fit one maximum-size request; Tier 1 allows 1,500 requests and 1 million tokens a minute; Tier 3 allows 15,000 requests and 10 million tokens a minute.
- Not supported: text generation, thinking, function calling, structured JSON output, search grounding, context caching and the Batch API.
It runs on the Gemini Developer API. Support in Google Cloud is listed as “coming soon”, and eesel notes it was not on Vertex AI as of November 12.
How it compares with OpenAI and others
Implicator reports that OpenAI’s decision model costs the same $0.10 per million input tokens, and that TypeSafe’s Jev 1.14 is cheaper at $0.042. eesel, which checked vendor pages in October, also lists Cloudflare’s Clef at $0.24 and says Microsoft’s model is $0.042. The sources disagree on when OpenAI’s endpoint left beta (eesel says early October, Implicator says general availability on October 22), so check OpenAI’s own page before comparing.
On quality, the evidence is mixed. Google says Decision 1 leads its benchmark suite with the lowest calibration error, and that it is nearly 5 times cheaper than Gemini 3.5 Flash-Lite on that suite. Implicator reports that in Google’s tests it led five of the seven tests all four models ran, with a calibration error of 0.0250 against Jev’s 0.0494. The independent Artificial Analysis Decision Index ranked it ninth of 18 at launch. On the AA-Paperclip test, which needs every one of 15 answers per document to be right, Gemini got all answers correct on 33% of documents, Jev on 35.5% and Nace AI’s Drex Max on 63%. For work where one wrong field matters, test on your own data first.
What businesses should check before using it
- Ambiguity: choice questions always spread probability across your options, even when none fits. Kingy relays Google’s advice to add an “other” option or run a separate yes/no check.
- Untrusted text: Google’s guide warns that user text can argue for its own classification, so high-stakes actions such as payouts need a human or second check.
- Safety filters: blocked requests return an error instead of answers.
- Unknowns: regions and data-handling terms were not stated on the pages we read, and the v1beta label suggests the API may still change.
Google names its own uses: routing queries in Gemini Notebook (it claims up to 40% lower latency for everyday queries), a safety layer for MCP tools in Google Search, and emoji reaction picking in CC from Google Labs. Pricing and terms can change, so confirm the current rates in Google AI Studio before you budget.
Sources
- Google: Gemini Decisions API (November 11)
- Implicator: Gemini Decisions API scores yes/no, choice and rubric questions
- Kingy: Gemini Decision costs, rate limits and limits
- eesel: Gemini Decisions API pricing and comparison
Image: Google headquarters in Mountain View, photo by Asoundd, CC BY-SA 4.0, via Wikimedia Commons.
Like our coverage? Add stepphase.com as a preferred source on Google to see more of our stories in Top Stories.
