Key details

  1. Released 11 November 2026; model gemini-decision-1.
  2. Up to 1,000 independent questions per call and 262,144 input tokens.
  3. Text, image, audio, video and PDF input; no generated text output.
  4. Yes/no, fixed-choice and rubric scoring with probabilities and confidence.
  5. Paid input $0.10 per million tokens; output zero billed tokens.
  6. Free tier available with different data-use terms.
  7. v1beta Decisions API; google-genai SDK 3.2+.
  8. No Batch API or context caching for this model.
  9. Google's 150–250ms latency is vendor-measured and should not be extrapolated to maximum input size.

What builders should take away

  1. Use decision endpoints for bounded routing and classification, not open-ended reasoning or generation.
  2. Evaluate calibration and false positives on labelled application-specific examples.
  3. Budget both shared input and question-instruction tokens; test real latency at expected payload sizes.
  4. Review free versus paid data-use terms and v1beta API compatibility before production adoption.

What changed

On 11 November 2026 Google launched Gemini Decision 1 (gemini-decision-1) through its Gemini Decisions API. Instead of generating prose, it scores up to 1,000 independent yes/no, categorical-choice and rubric questions against one shared multimodal input of up to 262,144 tokens. It returns probabilities and confidence scores. Supported inputs include text, images, audio, video and PDFs. The paid list price is $0.10 per million input tokens, with zero generated-output tokens billed; the free tier has separate data-use terms.

Why it matters

Agents often need a small decision—whether to call a tool, which specialist model to route to, or whether a document passes a check—not a verbose completion. A bounded API can reduce parsing and token overhead. Google's multimodal support and 1,000-question cap extend this pattern beyond text-only routing. But confidence scores require calibration against the application's own labelled cases, and low token pricing does not remove engineering, privacy or failure-handling costs.

One shared input, many independent checks

The API accepts a map of up to 1,000 questions over a shared input, returning typed outputs: noul (yes/no probability), choice among fixed options, or score against an ordered rubric. Questions are evaluated independently and cannot inspect one another's answers. The shared input is billed once, while each question adds instruction and criteria tokens.

Pricing and performance claims

Google's official pricing lists $0.10 per million paid input tokens, no output tokens, and no context caching or Batch API support. Google reports 150–250ms end-to-end latency, but its published low-latency measurements concern shorter contexts and smaller question counts than the maximum request. Real deployments should benchmark tail latency and calibration.

Integration boundary is a beta endpoint

Google's developer guide uses the Gemini Developer API v1beta endpoint and google-genai SDK version 3.2 or later. Developers must construct typed questions and handle probabilities, not treat the API as a replacement for arbitrary chat or free-form generation. Free-tier inputs can be used to improve Google's products, while paid-tier terms differ.

How it compares with the existing decision-model category

BTN already covers OpenAI's Decisions API in dossier #353. Google's announcement is distinct because of the multimodal input support and larger per-call question cap. Independent developer Simon Willison tested video input and observed non-extreme probabilities in a simple example. That is useful implementation evidence, not a comprehensive accuracy benchmark.

What to watch next

  • Independent accuracy and calibration studies against OpenAI and other decision APIs.
  • Latency and rate limits for large multimodal inputs and 1,000-question calls.
  • Stability of beta endpoint and availability on Google Cloud.

Still unclear

  • Published latency and benchmark advantages are primarily Google-produced.
  • Independent hands-on testing is limited and does not validate broad calibration.
  • Real token usage and reliability vary with modality and question construction.

Sources

Direct reading behind this dossier.

4 sources
Gemini Developer API pricing
Google AI Studio primary documentation

Current input pricing, free-tier data use and unsupported caching/batch.

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment