Google Introduces Gemini Decisions API and Gemini Decision 1 Model

What Happened

Google has announced the Gemini Decisions API, which uses the new Gemini Decision 1 model (gemini-decision-1) to perform fast, closed-set decisions [1]. This API is designed to evaluate up to 1,000 independent questions—including yes/no, multiple choice, and rubric-scoring tasks—in a single pass across text, images, audio, video, and PDFs [1].

Key technical details include:

• The model supports multimodal inputs up to 256K tokens [1].
• It provides end-to-end latency between 150-250 ms, with model compute latency between 50-100 ms [1].
• The API computes confidence scores from 0 to 1 for every answer, allowing developers to route ambiguous cases to humans or fallback systems [1].
• Pricing is set at $0.10 per 1M input tokens, with zero charges for output tokens [1].
• A no-cost tier is available in Google AI Studio for prototyping [1].
• The tool is described as being nearly 5x cheaper than Gemini 3.5 Flash-Lite using the Interactions API across a shared benchmark suite [1].

Integrations are rolling out for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with Gemini in Google Cloud support expected soon [1].

Separately, a bug report from a developer on the Google AI for Developers forum notes issues when migrating from the deprecated google.generativeai SDK (v0.8.6) to the new google.genai SDK (v1.68.0) while using gemini-3-flash-preview [2]. The report claims that the new SDK produces degraded multimodal output for tasks like OCR of 15th-century Spanish legal documents [2]. Specifically, the new SDK activates “thinking” tokens by default, which the developer states can increase character error rates from 50.5% to 88.8% and make processes approximately 3.5x slower [2]. The report also mentions that using thinking_level="MINIMAL" with a temperature of 0 can trigger repetition loops that generate up to 65,536 tokens [2].

Who Is Affected

• Developers building tools for model routing, policy checks, or content grading can use the new Gemini Decisions API [1].
• Users of the google.genai SDK may experience lower quality in multimodal transcription tasks compared to the deprecated google.generativeai SDK, which reached end-of-life in November 2025 [2].
• Teams using Stitch Rapid are already using the Decisions API to design product flows [1].

What Students Should Do Next

• Students and developers can access an API key through Google AI Studio to begin building [1].
• Documentation is available for those using Python, JavaScript, or REST [1].
• Developers can test the new model using demo apps such as Doorstep Parcels, Speed Sketch, or Find the Zoo Keeper [1].