Google's Gemini Decisions API scores up to 1,000 questions for 10 cents a million tokens

Google introduced the Gemini Decisions API on November 11, built on a new model called Gemini Decision 1, according to the Google AI Blog. The tool answers large batches of narrow questions in a single pass instead of writing out a full response, covering yes/no checks, multiple-choice picks, and rubric-style scoring.

According to Google, one request can carry a single multimodal input of text, images, audio, video, or PDF content up to 256,000 tokens, paired with as many as 1,000 separate questions. Because the model scores candidate answers directly rather than generating them word by word, Google says full responses return in 150 to 250 milliseconds including network time, with the model’s own compute taking roughly 50 to 100 milliseconds across loads of one to 100 questions.

Google prices the model at $0.10 per million input tokens and charges nothing for output tokens, since the API returns probabilities rather than generated text. A no-cost tier inside Google AI Studio lets developers test it for free. Google says this makes Gemini Decision 1 roughly five times cheaper than running the same workload through Gemini 3.5 Flash-Lite over its standard Interactions API. The company also says the model leads a suite of benchmarks covering agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error among the models it tested.

Google names several of its own teams already using the endpoint. Gemini Notebook uses it to route queries to the right model tier, which Google says cuts latency by up to 40% on everyday requests. The design tool Stitch uses it to turn product-flow sketches into playable wireframes within seconds. CC, a Google Labs project, uses it to pick a reaction from more than 200 emoji, and Google Search uses it to screen for adversarial prompts aimed at its tool-calling systems.

Simon Willison, writing on his own blog the same day, places the launch in a wider pattern: he says Google is now the third company with a dedicated “decision model,” after a product called Jev and a model from OpenAI. Willison says Gemini Decision 1’s three question formats mirror Jev’s almost exactly, down to reusing Jev’s “Noul” label for yes/no questions. On price, he notes Google’s $0.10 per million tokens matches OpenAI’s GPT-6 Luna decision model, while Jev stays cheapest at 4.2 cents per million. He also notes that Jev handles only text and JSON, OpenAI’s model adds images, and Gemini Decision 1 alone reads video and audio directly.

Willison tested the model against one of his own video files alongside yes/no questions. He reports that Jev and OpenAI’s model tend to answer with flat certainty, 1.0 or 0.0, while Gemini Decision 1 returned scores short of full confidence on comparable questions.

What this means

The two accounts agree on what the API does and what it costs. Google’s own post is the closer source on pricing, latency, and the internal use cases, since those figures come from the company’s own testing. Willison’s account adds what Google’s post leaves out: a direct comparison against Jev and OpenAI’s decision model, placing Gemini Decision 1 third to market even as Google claims it leads on calibration.

Unsettled is whether Gemini Decision 1’s less-certain scoring, which Willison observed in one informal test, reflects better calibration, as Google’s benchmarks claim, or just a different tuning choice. Neither report says how long the free Google AI Studio tier will last, or whether the 256,000-token input limit will rise.