Google has introduced the Gemini Decisions API, a new endpoint powered by the Gemini Decision 1 model that can evaluate up to 1,000 yes/no, multiple-choice and rubric-scoring questions in a single pass across text, images, audio, video and PDFs. Priced at $0.10 per million input tokens with no output token charges, the API achieves end-to-end latency of 150 to 250 milliseconds, according to the company.
The Decisions API is designed for bounded choices such as tool selection, policy checks, model routing and context grading, Google said. Rather than generating answer text token by token, Gemini Decision 1 scores candidate options in a single forward pass and returns probability and confidence scores.
Simon Willison said the release places Google alongside OpenAI and Jev in the decision-model category. Jev, the cheapest at 4.2 cents per million input tokens, handles only text and JSON, while OpenAI can handle text and images. Gemini Decision 1 supports audio, video and PDFs.
The model supports three question types: noul for yes/no checks, choice for selecting from 2 to 255 options, and score for ordered rubric grading. Every answer includes a confidence score between 0 and 1, allowing ambiguous cases to be routed to a fallback or human reviewer.
Google said internal teams are already using the API. Gemini Notebook employs it for dynamic model routing, Stitch for rapid design iteration, CC for emoji selection from a palette of over 200 options, and Google Search for adversarial behavior detection.
Simon Willison, who tested the model, reported that Gemini Decision 1 returned continuous probabilities rather than hard zero or one for yes/no questions, contrasting with competitors that return exactly 0 or 1. In his test with a video file, the model assigned a probability of 0.0498 to the presence of a dragon and 0.9234 to a robot, and estimated a 92.3% chance the video was set in the USA.
Gemini Decision 1 processes multimodal inputs of up to 256K tokens. The API is available through Google AI Studio with a no-cost tier for prototyping, and integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.
Analysis
Why This Matters
- The API enables developers to replace multiple specialised models or classifiers with a single multimodal pass, reducing both latency and cost.
- It introduces competition in decision-model pricing and capabilities, potentially driving wider adoption of decision-specific AI.
- Google's integration of audio and video into a decision model opens use cases in media moderation, multimodal inspection and voice-based routing.
Background
Decision models are a class of AI that compute probabilities for a set of predefined options in one forward pass, rather than generating free-form text. Jev was among the first, followed by OpenAI. Google’s entry adds full multimodal support to the category.
Key Perspectives
- Google: Says the Decisions API delivers fast, calibrated decisions across text, images, audio, video and PDFs, with confidence scores for routing and integration with major developer tools.
- Simon Willison: Reports that the model is competitively priced and fully multimodal, but notes that it returns continuous probabilities rather than discrete 0/1 answers, unlike some competitors.
What to Watch
- Adoption by the third-party platforms listed as integration partners.
- Any pricing adjustments from Jev and OpenAI as competition intensifies.
- Developer feedback on probability calibration in real-world applications.
Checked
12 claims in this story checked against 2 sources: 12 supported.
Show each claim
- SupportedGoogle introduced the Gemini Decisions API, powered by Gemini Decision 1.Google: AI (The Keyword): “Today, we’re introducing the Gemini Decisions API, a new endpoint for fast, closed-set decisions powered by our new Gemini Decision 1 model”
- SupportedThe release places Google alongside Jev and OpenAI in the decision-model category.Simon Willison's Weblog: “Google join Jev and OpenAI in the decision model category with "Gemini Decision 1".”
- SupportedThe API can evaluate up to 1,000 questions in one pass across text, images, audio, video and PDFs.Google: AI (The Keyword): “You can pass one multimodal input across text, images, audio, video, and PDFs (up to 256K tokens) along with up to 1,000 independent questions.”
- SupportedThe API returns results with 150–250 milliseconds of end-to-end latency, according to Google.Google: AI (The Keyword): “returning probabilities and confidence scores in 150-250 ms end-to-end latency at $0.10 per 1M input tokens with zero output tokens billed.”
- SupportedThe API costs $0.10 per million input tokens and has no output-token charges.Google: AI (The Keyword): “Gemini Decision 1 is priced at $0.10 per 1M input tokens and $0 for output tokens”
- SupportedJev is the cheapest of the compared models at 4.2 cents per million, and it handles text and JSON only; OpenAI handles text and images.Simon Willison's Weblog: “Jev remains the cheapest of the three, at 4.2 cents per million. Jev is text (and JSON) only. OpenAI can handle text and images.”
- SupportedThe API supports yes/no, multiple-choice questions with 2 to 255 choices, and ordered rubric scoring.Google: AI (The Keyword): “noul (Yes/No): Estimates the probability that a yes/no condition holds ... choice (Multiple Choice): Selects from 2 to 255 choices ... score (Ordered Rubric)”
- SupportedEvery answer includes a confidence score from 0 to 1.Google: AI (The Keyword): “Every answer includes a confidence score (from 0 to 1) so your code can route ambiguous cases to a fallback or human reviewer”
- SupportedGoogle says internal teams use the API for routing in Gemini Notebook, design iteration in Stitch, emoji selection in CC and adversarial-behavior detection in Search.Google: AI (The Keyword): “Teams across Google are already using the Decisions API as a low-latency control plane”
- SupportedSimon Willison reported that Gemini returned probabilities other than 0 or 1 in his yes/no video test, unlike Jev and OpenAI.Simon Willison's Weblog: “Jev and OpenAI both return 1.0 or 0.0 for yes/no questions, but Gemini insists on being a little less confident than that”
- SupportedIn Willison’s video test, Gemini assigned probabilities of 0.0498 to a dragon, 0.9234 to a robot and 92.3% to the video being set in the USA.Simon Willison's Weblog: “"dragon": { "type": "yes_no", "probability": 0.04976021841520312 }, "robot": { "type": "yes_no", "probability": 0.9234055387792485 } ... "usa": { "type": "yes_no", "probability": 0.9230459995157706 }”
- SupportedGemini Decision 1 supports multimodal inputs of up to 256K tokens.Google: AI (The Keyword): “You can pass one multimodal input across text, images, audio, video, and PDFs (up to 256K tokens)”
