Google launches Gemini Decision 1 judgment model and API

Google launched Gemini Decision 1, an AI model built for judgment and selection tasks rather than text generation, and the Gemini Decisions API on November 11, according to aitimes.com. The company announced the pair on its official blog.

The model returns probabilities and confidence scores between 0 and 1 instead of producing answers token by token. It supports true/false checks, multiple-choice selection with 2 to 255 options, and scoring against given criteria. A single input can carry up to 1,000 independent questions, processed in parallel across text, images, audio, video and PDFs, within a context of up to 256,000 tokens.

Google priced the model at $0.10 per million input tokens, with no charge for output tokens and input billed only once per input. The company says the cost is about one fifth of Gemini 3.5 Flash Light on the same benchmark setup. In a test with up to 16,000 input tokens and 100 questions, model computation took 50 to 100 ms and the API round-trip took 150 to 250 ms.

Google reported top scores in 7 of its 9 benchmarks and in 5 of 7 evaluations that included rival models. Its expected calibration error was 0.0250, against 0.0494 for Jev 1.14 and 0.0721 for GPT-6 Luna.

The API is already used inside Google products. Gemini Notebooks uses it to pick models in real time, cutting latency for ordinary queries by up to 40%, while Google Labs' CC uses it to choose among more than 200 emoji reactions, Stitch applies it to voice and touch interface edits, and Google Search uses it to detect malicious MCP tool requests.

Competitive context

OpenAI's Decision API and TypeSafe's Jev were already on the judgment-AI market before this launch. The OpenAI model behind its Decision API, GPT-6 Luna, is priced at $0.10 per million input tokens, while Jev 1.14 is priced at $0.042.