Gemini Decisions API: Google's Decision 1 Model for Routing, Policy and Scoring

Google launched the Gemini Decisions API and the Gemini Decision 1 model on November 11, 2026, giving builders a hosted way to score yes/no checks, multiple-choice routing, and ordered rubric grades in one multimodal pass instead of asking a chat model to write an answer. The announcement comes from Google's developer blog by Min Choi and Felix Yu, and it lands in the same month Sarvam, Hugging Face, and Microsoft pushed decision models for agent traces, guardrails, and routing.

TL;DR

Update (November 11, 2026): Decision 1 in AI Studio docs

Google's AI Studio Decisions API documentation now walks through yes/no, choice, and score questions against gemini-decision-1, matching the November 11 launch post. A same-day Hacker News submission titled Gemini Decision 1 is Free on Google AI Studio links those docs; it had no substantive comments when we checked—use it as a discovery link, not as measured community feedback.

If you prototype there, mirror production with Vertex billing rules before shipping agents—Studio quotas differ from Gemini 4 Carbon access paths we track separately.

Why decision APIs exist

Agent pipelines spend most of their wall-clock time on small judgements: pick a tool, allow or block a prompt, route a ticket, grade whether a step succeeded. A general chat model can answer those questions, but you pay for generated tokens, latency grows with decoding, and probabilities are often uncalibrated.

System One or decision models flip the contract: you supply a state (the multimodal input), a question, and for choice types a fixed candidate set. The model returns probabilities so code can branch above a threshold and escalate ambiguous cases to a human or a larger model.

explainx.ai tracked that pattern when Sarvam released Prism, when SubQ shipped Snap for long-context grading, and when Hugging Face documented System One request shapes. Google's entry is significant because it bundles native multimodal inputs, a first-party endpoint, and zero output-token billing at a headline price many routers already list.

What Google actually shipped

Google describes Gemini Decision 1 as purpose-built for closed-set decisions: tool selection, policy checks, model routing, and context grading. Instead of scoring options while emitting answer text token by token, the model runs a single forward pass and returns structured probabilities.

Three question primitives

Google names three decision types you can mix in one request:

Each answer includes a confidence score from 0 to 1, which Google positions for fallback routing when confidence is low.

One input, many questions

The API is built around a shared input blob: a paragraph, a call transcript, a PDF claim, or a bundle of modalities. You attach a questions object with your own keys (is_damaged, support_queue, and so on). Google stresses that input tokens are charged once no matter how many questions you hang off that state; only per-question instructions and criteria add tokens.

That design matches how teams already batch evaluations in agent harnesses, but collapses them into one network round trip when latency matters.

Multimodal in one pass

Google says Decision 1 inherits Gemini's native multimodal stack: text, images, audio, video, and PDFs without a separate OCR or transcription line item. The post gives composite examples such as pairing a call recording with account metadata or a scanned claim with damage photos, with resolution knobs on visual inputs to trade detail against token cost.

For builders comparing vendors, note that third-party gateways may roll out modality support at different speeds. Vercel's changelog for November 11, 2026 says PDF, audio, and video require AI SDK 7, while text and image paths work on older TypeSafe-compatible routes.

Latency, benchmarks, and cost claims

Google's post includes vendor benchmark charts (we could not mirror the image assets through this environment's fetcher, so treat figures as Google-reported until independent testers publish runs).

Headline claims from the announcement:

• 50–100 ms median model compute, 150–250 ms end-to-end latency.
• Latency stays nearly flat from 1 to 100 questions on contexts up to 16K tokens (per Google's chart narrative).
• Against Gemini 3.5 Flash-Lite with the Interactions API on a shared internal suite, Google says Decision 1 is nearly 5x cheaper because output tokens are not billed.

Pricing on launch day: $0.10 per million input tokens and $0 for output tokens, plus a no-cost tier in AI Studio for prototyping.

Provider pages echo the same list price. OpenRouter's model card lists 262,144 context tokens and November 11, 2026 as the release date, which helps if you route through a broker instead of Google direct.

How to call it (REST shape)

Google documents REST, Python, and JavaScript guides on the announcement page. The REST skeleton third-party write-ups repeat looks like this:

curl -X POST "https://generativelanguage.googleapis.com/v1beta/decisions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-decision-1",
    "input": "The blue mug I ordered on Monday arrived cracked on Wednesday.",
    "questions": {
      "is_damaged": {
        "type": "noul",
        "instructions": "Did the item arrive damaged?"
      },
      "queue": {
        "type": "choice",
        "instructions": "Which support queue should handle this message?",
        "criteria": {
          "billing": "Payment, invoice, or refund questions",
          "replacements": "Damaged, missing, or wrong items",
          "general": "Product and account questions"
        }
      }
    }
  }'

Two practical details trip teams during early integrations:

• Authentication uses x-goog-api-key, not a Bearer header on Google's first-party endpoint.
• Question keys are yours; responses use the same keys, which keeps agent orchestration code readable.

If you already use the Gemini API on Vertex, expect separate rollout notes; the November 11 post centers Google AI Studio and the Gemini API developer path.

Demos Google wants you to try

Google links three AI Studio demo apps on the launch post: Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. They are marketing-friendly, but they signal intended UX: human-visible confidence, fast iteration on rubrics, and multimodal inputs rather than long chat threads.

For agent builders, the more relevant comparison is internal: can you replace three chained generateContent calls (classify, route, score) with one decisions request and explicit thresholds?

Comparison table: hosted Gemini vs other decision models

Google is not claiming the highest accuracy on every public decision index Sarvam plotted. The pitch is multimodal breadth, predictable closed-set APIs, and no output-token surprise bills on high-volume routing.

Where this fits agent architecture

Decision endpoints belong between retrieval and action:

1. Ingest multimodal state once (ticket, call, PDF, screen recording).
2. Decide with thresholds (allow tool call, pick model tier, escalate).
3. Act with a worker model or human only when confidence clears the bar.

That pattern shows up in Raindrop's memory routing, in Amazon Bedrock AgentCore evaluation, and in Claude Code routing guides. Gemini Decision 1 does not replace those systems; it gives teams on Google's stack a native primitive instead of prompt-hacking chat completions.

Policy and abuse teams should still pair any classifier with human review paths and with platform rules. Retailers are already drawing lines on who may act on behalf of users, as in eBay's third-party agent limits and United's November 2026 bot booking ban.

Calibration and testing discipline

Google emphasizes confidence scores on every answer. That only helps if you plot calibration curves on your own traffic: at 0.9 stated confidence, how often is the model wrong on your rubric?

Sarvam's Prism launch included an explicit expected versus actual error table on high-confidence answers. Google has not yet published an equivalent public table for Decision 1 in the text we fetched. Until third parties replicate, run a shadow period: log decisions, do not auto-act, compare against human labels for two weeks.

For long-context document QA, also test whether 256K context marketing matches your PDF tokenization bills. SubQ's Snap post is a useful reference for how per-million pricing interacts with huge states when every routing call re-embeds the document.

Migration notes for existing Gemini apps

Teams with generateContent classifiers should plan a separate code path:

• Swap model IDs is not enough; request and response schemas differ.
• SDK support may use client.decisions.create(...) in Python per integrator docs.
• Brokers like OpenRouter and Vercel expose /v1/decisions with their own auth headers.

Keep a fallback chat model for open-ended explanations. Decision APIs answer structured questions; they are the wrong tool for user-facing prose.

What's still unverified

Independent benchmark runs on Gemini Decision 1 were not available at publish time on Hacker News or GitHub. Google's 5x cheaper and flat latency to 100 questions claims come from Google's benchmark suite against Gemini 3.5 Flash-Lite, not from Artificial Analysis or community harnesses.

Treat launch-week numbers as vendor baselines. Replicate on your prompts, especially if you mix low-resolution video with dense PDF tables, before you wire auto-purchase or auto-refund workflows.

What to do this week

1. Open AI Studio, enable the Decisions API, and port one production routing rubric (support queue or model tier).
2. Log confidence histograms for a week before enabling automatic actions.
3. Compare spend against your current chat-classifier path using the same traffic; include embedding and OCR costs you may drop.
4. If you need on-prem or Apache 2.0 weights, benchmark against Sarvam Prism on identical questions.
5. Read Google's Doorstep Parcels demo for UX patterns on showing probabilities to end users.

Related reading

• Sarvam Prism: open-weights decision model
• Hugging Face System One decision models hub
• SubQ Snap long-context decision model
• How to wire Jev into agent routing
• Gemini 4 Carbon on Vertex access
• eBay blocks third-party shopping agents
• OpenAI and Singapore Airlines agentic travel APIs

Sources: Google launch post (Nov 11, 2026) · OpenRouter model card · Vercel AI Gateway changelog