Google Launches Decision API at $0.10 per Million Tokens

Google offers a decision API that directly returns calibrated probabilities instead of text. Announced at $0.10 per million input tokens and reported to have an end-to-end latency of 150 to 250 ms, it arrives without caching or batching and remains more expensive per token than Jev.

No caching or batching, and a higher entry cost than Jev

The API currently does not offer prompt caching or batch API, meaning that any repeated context is fully charged again. At $0.10 per million input tokens, Gemini Decision 1 is still more expensive than Jev 1.14, which is priced at $0.042 per million. In the comparison table, caching is indicated as unsupported on Google's side, supported on OpenAI Decisions, and undisclosed for Jev. In terms of latency, Google reports 150 to 250 ms end-to-end, while OpenAI Decisions is described as being about ten times faster than its API Responses, with no equivalent figure published for Jev.

Internal uses and ecosystem deployment

Google reports internal usage in Gemini Notebook to route to different model levels, achieving up to 40% lower latency on daily queries. Google Labs utilizes Creative Chips to select emoji reactions from over 200 possibilities. The API is used by Stitch's rapid canvas to generate playable mockups in seconds. Support in the ecosystem is currently being integrated into OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with support in Gemini on Google Cloud announced as forthcoming.

A multimodal decision API that returns calibrated scores

Google launches Gemini Decisions, based on the specialized model Gemini Decision 1, which does not generate text and directly returns calibrated probabilities. The model is multimodal and processes text, images, audio, video, and PDFs in a single pass, with responses accompanied by a confidence score between 0 and 1. Each request includes a model ID, shared input, and a question map, with these evaluated in parallel and without mutual interference. Three question formats are planned: yes/no with a bounded probability between 0.01 and 0.99, single choice from 2 to 255 options with probabilities per option, and scoring on 2 to 255 ordered levels. The API handles up to 1,000 questions per call with 0 output tokens. Google recommends redirecting low-confidence cases to a fallback plan or human review.

Announced latency and billing focused on shared input

Google reports a median computation time between 50 and 100 ms and an end-to-end latency ranging from 150 to 250 ms, these values applying to inputs up to 16K tokens and remaining nearly stable for a number of questions ranging from 1 to 100. The shared input is charged only once per request, with each question incurring fees only for its own instruction and criteria tokens, with no additional cost for reflection, transcription, or OCR tokens. The available context capacity is 262,144 tokens for shared input and questions. The service is accessible only via the Gemini Developer API in v1beta, billed at $0.10 per million input tokens, and supports up to 1,000 questions per request.

Self-reported benchmarks and measured calibration

According to figures provided by Google, Gemini Decision 1 achieves the best performance on 7 of the 9 published accuracy benchmarks, with the lowest calibration error. The model scores 92.2% on JevBench and records its lowest Macro-F1 at 58.7% on Amazon ESCI. It stands at 79.9% on When2Call compared to 82.0% for Jev 1.14. Other reported scores include 79.7% on New Yorker Caption, 86.2% on Mind2Web Multimodal, 79.8% on PubMedQA, 86.0% on Global-MMLU-Lite, 71.8% on MMAU, and 69.1% on Video-MME without subtitles. The Top-1 ECE error is reported at 0.0250, about half that of Jev 1.14 at 0.0494. For seven shared evaluations, the total cited price is $0.88 for Gemini Decision 1 compared to $0.68 for Jev 1.14, $0.93 for GPT-6 Luna, and $4.14 for Gemini 3.5 Flash-Lite. Google further claims that Gemini Decision 1 is nearly five times cheaper than Flash-Lite on this suite.

Developer access and large-scale comparisons

The endpoint is active on the Gemini Developer API v1beta and requires the google-genai SDK 3.2 or higher, with a no-cost tier for prototyping in Google AI Studio and guides for Python, JavaScript, and REST. In published comparisons, OpenAI Decisions shows a context of 1,050,000 tokens on its model page, while Jev 1.14 indicates 64K per request, extendable to 262K with opt-in header. Gemini Decision 1 accepts from 1 to 1,000 questions per request compared to 1 to 400 for OpenAI Decisions, and Jev is limited by context without a specified ceiling. The listed input modalities go up to PDF, video, and audio for Gemini Decision 1, while OpenAI Decisions cites text, image, and JSON, and Jev only text. The referenced input prices are $0.10 per million for Google, $0.10 non-cached for OpenAI, and $0.042 for Jev. The Google API remains limited to Google hosting, with no self-hosting option.