AI · · 5 min read

Google's Gemini Decisions API Scores Up to 1,000 Questions in One Pass, with No Output Tokens Billed

The new gemini-decision-1 model doesn't generate text. It scores yes/no, multiple-choice and rubric questions against one multimodal input in about 150–250 ms, for $0.10 per million input tokens. Google pitches it for model routing, tool selection and policy checks.

Ink-and-gouache illustration of a single dark line on cream paper that fans out into five thin branches, each ending in a dot of a different size, the largest one ochre in the middle, above a sage strip
Illustration by TechPulse Editorial · One input, several options, and a weight on each.

A lot of what agent harnesses ask an LLM isn’t open-ended. Which tool should run next? Does this request need the big model? Does this message break a policy? Teams usually answer these by calling a small generative model and parsing its text, or by training a classifier. On November 11, 2026, Google shipped a third option: the Gemini Decisions API, a separate endpoint for closed-set decisions, backed by a new model, Gemini Decision 1 (gemini-decision-1).

How it works

Gemini Decision 1 doesn’t decode an answer token by token. It scores the candidate options in a single forward pass and returns probabilities. A request has two parts:

  • One shared input: text, images, audio, video and PDFs, mixed freely, up to 256K tokens. It uses the same Content block format as the Interactions API.
  • Up to 1,000 independent questions against that input, each one of three types:
    • noul: probability that a yes/no condition holds, with optional criteria for true and false.
    • choice: picks from 2 to 255 options and returns normalized probabilities for every candidate.
    • score: grades against an ordered rubric starting at 0, returning a weighted score and per-level probabilities.

Every answer includes a confidence score from 0 to 1, so your code can send ambiguous cases to a fallback model or a human.

Speed and price

Google quotes median model compute of 50–100 ms and 150–250 ms end to end for inputs up to 16K tokens, staying nearly flat from 1 to 100 questions per request. Pricing:

  • $0.10 per 1M input tokens, $0 for output, no thinking-token charges and no separate audio transcription or OCR fees.
  • The shared input is billed once per request, however many questions you ask. Each question adds only its own instruction and criteria tokens.
  • A no-cost tier in Google AI Studio for prototyping.

Google says it leads its benchmark suite (agent routing, reasoning, ranking, multimodal understanding) with the lowest calibration error. It also says it’s nearly 5x cheaper across that suite than Gemini 3.5 Flash-Lite through the Interactions API. These are Google’s own numbers.

What Google uses it for

  • Model routing in Gemini Notebook: classifying intent to send each query to the right model tier and thinking level, which Google says cut latency by up to 40% on everyday queries.
  • A safety layer for MCP tools in Search, detecting adversarial behavior within the latency budget.
  • Lighter uses in Stitch (applying voice and tap edits to wireframes) and CC (picking one of 200+ emoji reactions).

Integrations are rolling out for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs. Support in Gemini on Google Cloud is listed as coming soon.

Why it matters for agent builders

The launch landed a day after Google opened Gemini 4 Carbon to all paying developers and previewed /swarm in Antigravity, where Carbon handles coordination and auditing while lighter models do the parallel work. Decisions is the cheap piece in that kind of setup: the router or gate that decides which model, tool or reviewer gets a task.

In practice:

  • Routing and gating become a scoring call, not a generation call. You get calibrated probabilities instead of parsing “yes” out of a sentence, and you can set thresholds per question.
  • Batch your checks. One screenshot or document, 20 policy and quality questions, one request, with the input billed once.
  • Keep generative models for open-ended work. Decisions only chooses from options you define. It doesn’t explain or write anything, so it complements your main model rather than replacing it.
  • Measure before trusting the calibration claims. Run it on a labeled slice of your own routing or moderation traffic and compare against what you use today.
from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})

result = client.decisions.create(
    model="gemini-decision-1",
    input=[{"type": "text", "text": user_request}],
    questions={
        "tier": {
            "type": "choice",
            "instructions": "Which model tier should handle this coding request?",
            "criteria": {
                "fast": "Small edit, lookup or explanation.",
                "frontier": "Multi-file change, debugging or design decision.",
            },
        },
    },
)

Sources & Primary Advisories

  1. Google Blog — Gemini Decisions API: Fast, calibrated multimodal decisions (Nov 11, 2026)
  2. Google Blog — Gemini 4 Carbon is now available (Nov 10, 2026)
Share Dispatch