News · 2026-11-12

Google's Gemini Decisions API picks an answer in one pass and writes no text

Google DeepMind released the Gemini Decisions API on November 11, 2026, built around a model called gemini-decision-1 that picks among options you supply instead of writing an answer. According to Google's blog and docs, it scores every candidate in a single forward pass, produces no output tokens, and costs $0.10 per million input tokens with nothing charged for output. Google says end-to-end latency is 150 to 250 milliseconds.

Key facts

• Who and when: Google DeepMind, announced November 11, 2026 by product manager Min Choi and principal scientist Felix Yu; call shape is client.decisions.create(...) with api_version="v1beta" and Python SDK 3.2 or later.
• Headline number: $0.10 per 1M input tokens and $0 for output, with a no-cost tier in AI Studio (Google's figures).
• Limits: up to 262,144 shared input tokens, 1 to 1,000 questions per request.
• Primary source: Google's Decisions API docs and the gemini-decision-1 model page.

Why a model that does not write

Most developers use a large language model for yes/no and routing work the awkward way: ask it a question, let it write a sentence, then write code to parse the sentence. That costs many sequential decoding steps and brittle parsing. The Decisions API removes both. In Google's own words, "Rather than generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass." The model page calls it a single prefill pass, the same step a normal model does to read your prompt, with the writing stage skipped entirely. If you want the background, our lesson on discriminative models explains why choosing from a closed set is so much cheaper than composing free text.

Think of it as the difference between asking a clerk to write you a memo and asking them to tick one box on a form. The box is faster, cheaper, and trivial for a program to read.

What you can ask

There are three question types. A yes/no question returns a probability clamped between 0.01 and 0.99, so a threshold above 0.99 never fires. A choice question takes 2 to 255 options and returns the winner plus a normalized probability for every option. A score question is an ordered rubric of 2 to 255 levels (Google recommends 2 to 10) and returns a weighted score plus per-level probabilities. Every answer carries a 0-1 confidence value, which Google's blog says code can use to "route ambiguous cases to a fallback or human reviewer".

Because the input is shared, you pay for it once per request and each extra question adds only its own instruction tokens. Google says latency stays nearly flat from 1 to 100 questions, measured on contexts up to 16,000 tokens. The model never sees question keys, so the full question must live in the instructions field.

Who is already using it

Google names four of its own products. Gemini Notebook uses it as a real-time router to pick a model tier and thinking level, for what Google calls "up to 40% lower latency" on everyday queries. Stitch uses it to generate playable wireframes in seconds, a Labs product called CC uses it to pick from more than 200 emoji reactions, and Google Search uses it as a safety layer on MCP tools to detect adversarial behavior. Outside integrations are rolling out through OpenRouter, LangChain, the Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.

How it compares with rivals

This is a crowded category. TypeSafe's Jev, Microsoft and OpenAI all ship decision-style models; see our coverage of Jev, Microsoft's Decision 1 and OpenAI's Decisions API. Latent Space's AINews issue "D-Day: Decision Models Day" led with the launch. By that newsletter's account, Artificial Analysis gave Decision 1 an index score tying Jev 1.14 and trailing two other models by seven points, and an average of 211 milliseconds per task against 92 for Jev. Those figures are second-hand from the newsletter, not from the Artificial Analysis page itself. Google's "nearly 5x cheaper" claim is against Gemini 3.5 Flash-Lite, a generative baseline, so do not mix it with per-token prices from other vendors. The wider Hacker News thread had about 458 points when we checked.

How practitioners are using the category

Latent Space framed these as "System One" models: "non-generative models that return structured choices, scores, or probabilities", typically "much faster and cheaper than reasoning LLMs". At the AI Engineer Code Summit, Glean said it uses them next to LLMs because "As the taxonomy of the decision space changes, we don't have to fine-tune these models again and again." TypeSafe's Diogo Almeida put the underlying problem more bluntly: "you get Einstein 95% of the time, but Mr. Bean 5% of the time".

The honest caveats

Google's docs are candid about the failure modes. A choice question spreads probability across your options even when none fits, so add an "other" key (not a positional catch-all like "none_of_the_above", since exact ties go to the alphabetically first key) or check confidence. Score averages can hide the shape of the distribution. Unrelated content in the input lowers accuracy. And the input is untrusted: "User-supplied text can argue for its own classification", so high-stakes actions such as payouts need a second check. Google has not published how the model produces its probabilities, so any architecture talk is speculation, and the "lowest calibration error" claim comes from charts we could not read as text.

One more practical limit: context caching, the Batch API and Flex or Priority inference are all unsupported, so a long document that many questions refer to across separate requests is billed in full each time. Packing all questions into one request is the intended pattern.

Primary source, verified: read the paper →

Ground Truth. (2026, November 12). Google's Gemini Decisions API picks an answer in one pass and writes no text. Ground Truth. https://groundtruth.day/news/google-gemini-decisions-api-scores-answers-without-writing-text.html

@misc{groundtruth:google-gemini-decisions-api-scores-answers-without-writing-text,
  title  = {Google's Gemini Decisions API picks an answer in one pass and writes no text},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/google-gemini-decisions-api-scores-answers-without-writing-text.html}
}

Topics: google · gemini · decision-models · api · inference · pricing