Google on November 12 2026 introduced Gemini Decision 1 and the Gemini Decisions API, a new class of model that does not generate text token by token. Instead, it scores up to 1000 closed-set candidate answers in a single forward pass and returns probabilities with confidence scores.
A Model That Decides, Not Generates
For tasks that only need a verdict - spam or not, which agent should handle a task, how a document should be classified - this removes the entire text-generation step. A developer supplies the input, the question, and a fixed set of candidate options, and the model returns the chosen answer, its probability, and a confidence score.
Benchmarks
On the Artificial Analysis Decision Index v2.1, Gemini Decision 1 scores 66, tying GPT-6 Luna (Max) at 66 and ahead of d1-ultra at 64 and GLiDE v1.2 at 61. Across nine internal benchmark categories it took first place in seven, with a top accuracy of 92.2 percent.
Key points
- 0.10 dollars per 1 million input tokens (output tokens free)
- 150 to 250 milliseconds latency per decision
- 66 on the Artificial Analysis Decision Index v2.1
- 1000 questions scored in a single forward pass
- 255 maximum multiple-choice options
- 5 times lower cost than Gemini 3.5 Flash-Lite
Read the full report on ThreeRouter AI Daily: https://www.threerouter.com/blog/414
