- Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1.
- The API scores up to 1,000 yes/no, multiple-choice, or rubric questions against one shared input and returns calibrated probabilities instead of generated text.
- Pricing is $0.10 per million input tokens, with zero output tokens billed, and it runs only on Google’s hosted Gemini Developer API.
- Google says Decision 1 leads 7 of 9 accuracy benchmarks with the lowest calibration error among models tested.
What Happened
Google launched the Gemini Decisions API on November 11, 2026, powered by a new model called Gemini Decision 1, Marktechpost reported the same day. Unlike a standard chat model, the API does not generate text — it scores up to 1,000 typed questions against one shared input across text, images, audio, video, and PDFs, returning calibrated probabilities.
Why It Matters
The launch targets a specific pain point inside AI agents: the small branching decisions — routing, policy checks, tool selection, grading — that agent frameworks currently handle with full chat models. Google’s move follows a similar bet by OpenAI’s Decisions API, built around the Jev model, which targets the same agent-decision bottleneck.
Technical Details
Gemini Decision 1 supports a shared context window of 262,144 tokens across the input and all questions, runs only through Google’s hosted v1beta Gemini Developer API with no open weights or self-hosting option, and is priced at $0.10 per million input tokens with zero output tokens billed. Google’s own benchmark table shows Decision 1 leading on 7 of 9 accuracy tests, scoring 92.2% on the public split of JevBench but only 58.7% Macro-F1 on Amazon ESCI, its weakest result.
Who’s Affected
Developers building agent frameworks that need fast, cheap classification calls are the primary audience, since the API is designed to replace ad hoc chat-model calls for routing and grading. Google is directly competing with Jev 1.14, which beat Decision 1 on the When2Call benchmark, 82.0% versus 79.9%.
What’s Next
Decision 1 is not the cheapest option in its category: benchmarked on the same question suite, Jev 1.14 cost $0.68 compared with Decision 1’s $0.88. Teams choosing between the two will need to weigh Google’s accuracy lead against that per-call cost gap.
