Google Prices a Model That Only Picks Answers

MOUNTAIN VIEW, Calif., Nov. 11 — GoogleGoogle on Wednesday put a price on a model that does not write. Gemini Decision 1, sold through a new Decisions API, takes a question and a fixed set of answers and returns a probability for each, at 10 cents per million input tokens and nothing for output, because it generates none.

One request can put up to 1,000 yes-or-no, multiple-choice or rubric questions to the same text, images, audio, video or PDFs. Google reports a median of 150 to 250 milliseconds end to end for contexts of up to “16K tokens.”

The work it is priced for is the plumbing of AI agents, what Google calls “bounded choices like tool selection, policy checks, model routing, and context grading.” Most teams, The New Stack wrote in September, handle them with a chat model and “a carefully worded prompt.” On input, Decision 1 costs what Anthropic charges for prompts of up to 100,000 tokens to Claude Haiku 5.5, a model it lists for “classification, extraction, and routing”; on output, at that prompt length, Haiku costs 50 cents per million tokens and Decision 1 nothing. Google lists its own Gemini 3.5 Flash-Lite at 30 cents for input and $2.50 for output.

Google says that, across its own benchmark suite, Decision 1 is “nearly 5x cheaper” than Flash-Lite run through its generative endpoint, and that it “leads across the suite, with the lowest calibration error.” Those are Google’s figures, on Google’s tests. Its own products already use it, Google said, among them Gemini Notebook, to route queries.The RouterGoogle says Gemini Notebook uses the API to send each query to a model tier and thinking level, serving everyday queries with “up to 40% lower latency” while reserving deep reasoning models for complex tasks.

Google is not first: The New Stack reported that a start-up, TypeSafe, introduced a decision model called Jev in September, and that OpenAI announced a Decisions API, built on its small GPT-6 Luna model, on Sept. 29. By Oct. 9, The New Stack counted OpenAI, Upstage, Perplexity and Cloudflare among the providers of hosted decision models, with OpenAI’s still in limited preview and without published pricing. Google’s three question types, noul (yes or no), choice and score, carry the names The New Stack reported for TypeSafe’s.

Google’s documentation carries its own warnings. “Treat input as untrusted,” it says. “User-supplied text can argue for its own classification.” It advises keeping “a human or a second check on high-stakes actions such as payouts,” and notes that a choice question spreads its probability across the options given “even when none fits.” The API supports neither batch processing nor context caching, so repeated input is billed in full each time; Anthropic sells Haiku’s input at half price in batches.

The launch comes in a busy week for prices. On Monday Anthropic cut its flagship’s price by a fifth, to $8 per million input tokens and $40 per million output. At the top of Google’s own list the move runs the other way: on Dec. 1, Gemini 4 Argon’s introductory rates of $2 and $10 end and it goes to $4 and $20, as Google set out in October.

Decision 1 is available through the Gemini API in Google AI Studio, with a free tier for prototyping; support in Google Cloud is “coming soon,” Google said.

The same day, Google DeepMind released AlphaGenome 2, which predicts how DNA variations affect specific human cell types. DeepMind says it identifies “nearly twice as many variants that change gene expression in cells” as its predecessor. It is open to non-commercial research through the AlphaGenome API and in private preview for commercial research, and is not approved for clinical or diagnostic use.