Google launches Gemini Decisions API for fast multimodal choices

What Google released

Google announced the Gemini Decisions API on November 11, 2026, introducing gemini-decision-1 for closed-set judgments through the Gemini API and Google AI Studio. The launch price is $0.10 per million input tokens, with zero output tokens billed and a no-cost tier for prototyping. Google says one request can evaluate up to 1,000 independent questions against a shared multimodal input of up to 256K tokens.

This is not a new chat model that writes a shorter answer. Google describes a model that scores supplied options in a single forward pass, returning typed answers, probabilities and confidence rather than generating prose token by token. The intended jobs include choosing a model or tool, checking a policy condition and grading retrieved context.

The API documentation limits availability to the Gemini Developer API v1beta with Gemini Decision 1. Google’s announcement says ecosystem integrations are rolling out and Gemini in Google Cloud support is coming soon. Do not assume an existing Cloud deployment or third-party SDK already exposes the new endpoint.

Three question types, not an open-ended answer

The yes/no type is named noul. It returns the probability that a condition holds, constrained to 0.01 through 0.99. That matters when writing thresholds: a requirement above 0.99 can never match. Confidence for this type is the larger of the yes and no probabilities; it is not another independent assessment.

The choice type selects among 2–255 options and returns normalized probabilities. It will distribute probability across the supplied choices even if none is appropriate. Google recommends a scope check, confidence threshold or explicit other option. Option keys are sorted alphabetically in UTF-8 byte order, with exact ties going to the alphabetically first key—not the first option in your JSON.

The score type uses an ordered rubric indexed from zero. Its result is a probability-weighted average of level indices, not necessarily a discrete category. The documentation permits 2–255 levels and recommends 2–10 for most rubrics. Two different distributions can have the same average, so inspect per-level probabilities when an ambiguous middle score would hide a consequential disagreement.

The latency claim is narrower than the request ceiling

Google reports 50–100 ms median model compute and 150–250 ms end-to-end latency across multimodal inputs with contexts up to 16K tokens. It describes latency as nearly flat from 1 to 100 questions per request. Those are vendor-reported results, not AccessAllGPT measurements, and neither range establishes latency at 256K tokens or 1,000 questions.

Google also claims the lowest calibration error across its comparison suite and nearly five times lower cost than Gemini 3.5 Flash-Lite through the Interactions API on the shared benchmarks. We did not reproduce those comparisons. Calibration on a vendor benchmark does not establish that a 0.9 confidence value is reliable on your company’s documents, accents, screenshots or unusual support tickets.

The useful change for builders is the interface: ordinary code can consume probabilities without parsing generated text. Whether that also improves speed, reliability or total cost in a particular workflow remains an evaluation question. Include network time, media processing, retries and escalations in that comparison.

Shared-input billing helps repeated checks, not repeated requests

Google says shared input is billed once per request regardless of the number of questions, while every question adds its instruction and criteria tokens. The announcement lists no output-token charges, variable thinking-token bills or separate audio-transcription and OCR costs. These statements do not make media input free: input token usage still matters.

The documentation explicitly excludes context caching, Batch API, Flex and Priority inference. Repeating a document in a second request counts the input again in full. Group genuinely independent checks when possible, but do not bundle unrelated questions simply to reach the maximum.

Questions cannot see each other’s answers. If a second decision depends on the first, either make another request after reading the result or ask both against the shared input and ignore the inapplicable answer in code. A thousand questions in one request is not a thousand-step reasoning chain.

Multimodal input comes with integration constraints

The API accepts text, images, audio, video and PDF blocks using the Interactions API content format, but question instructions are text-only. Tool-call, tool-result and thought blocks are unsupported, and conversation context must be serialized as text. An agent transcript is therefore not automatically a valid Decisions API input.

Google’s documentation allows inline data, Files API URIs and public or signed HTTPS URLs, but not YouTube URLs or direct gs:// Cloud Storage URIs. It lists up to 100 MB inline per request, or 50 MB for PDFs, and up to 15 MB per external URL. The MIME type must match the actual format. Adjacent text blocks are concatenated without separators, so supply your own spaces or newlines.

Safety-filtered requests return HTTP 400 rather than answers. Applications need an explicit error path as well as a low-confidence path. A classifier result is also not authorization: user input can argue for its own category, and Google recommends adversarial examples and human or second checks for high-stakes actions.

What builders should try next

Start with one bounded decision that currently uses a generative call: routing a public support example, choosing a tool from a short allowlist or scoring a retrieved passage against a clear relevance rubric. The documentation requires google-genai SDK 3.2 or later; verify the relevant language package and endpoint before changing a working integration.

Compare against the existing route on labeled examples, including out-of-scope inputs. Set confidence thresholds from false-positive and false-negative costs rather than picking a reassuring number. Keep deterministic arithmetic, date comparisons and permissions in ordinary code; use the model for the judgment it is actually being asked to make.

AccessAllGPT did not call Gemini Decision 1 or test Google’s no-cost tier. This is a source-backed launch briefing, not a hands-on review. The immediate recommendation is a reversible trial with an explicit fallback—not replacement of a production router based on headline latency alone.

Disclosures

AccessAllGPT received no payment, credits or private vendor briefing for this article. Google did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google. Publication-wide relationships are listed on the disclosures page.