Google Introduces Gemini Decisions API for Focused AI Answers

Google's Gemini Decisions API evaluates up to 1,000 bounded questions against shared multimodal inputs, returning probabilities and confidence scores for

Quick Answer

Google’s Gemini Decisions API gives developers a way to evaluate up to 1,000 yes-or-no, multiple-choice or scoring questions against shared text, image, audio, video and PDF inputs. Each answer includes a probability and a confidence score, making the endpoint suited to policy checks, content grading, tool selection and model routing rather than open-ended answer writing. Google announced the service on November 11, 2026 and listed the paid API at $0.10 per 1 million input tokens with no output-token charges.

For business owners, the practical distinction is to consider how structured results may fit into an approval rule, review queue or routing path. The model is still experimental, and its reported performance comes from Google-defined benchmarks, so a business-specific evaluation remains necessary before operational use.

What Is the Gemini Decisions API?

The Gemini Decisions API is an endpoint powered by Gemini Decision 1 that scores predefined options in a single forward pass. Google describes three question formats: yes-or-no conditions, multiple-choice selections and ordered-rubric scoring.

  • Yes-or-no: The model estimates the probability that a specified condition applies.
  • Multiple choice: The model selects from 2 to 255 options and returns normalized probabilities for each candidate.
  • Ordered rubric: The model applies a user-defined scale and returns a continuous weighted score with probabilities for each level.

Every answer has a confidence score from 0 to 1. Google documents that score as a way to route ambiguous cases to a fallback or human reviewer. A single request can combine different question types and share the same multimodal input, such as a call recording paired with account metadata or a scanned claim paired with photographs of damage.

The shared input can contain up to 256,000 tokens, while one request can contain up to 1,000 independent questions. According to AlphaLab’s November 13, 2026 analysis, the combined input and questions fall within a 262,144-token context window, and shared input is billed once per request. Each question still adds instruction and criteria tokens.

How Does It Differ From a General AI Model?

General language models produce free-form text that an application may need to parse, check and filter. Gemini Decision 1 instead scores specified options without token-by-token answer generation and returns the selected option, probability and confidence in a defined structure.

Google DeepMind product manager Min Choi and principal scientist Felix Yu reported median model-compute latency of 50 to 100 milliseconds and end-to-end latency of 150 to 250 milliseconds. In the November 11, 2026 Google announcement, those measurements cover multimodal inputs and contexts up to 16,000 tokens, with latency described as nearly flat from one to 100 questions per request. The 256,000-token input capacity and 1,000-question request limit are separate limits from that measured latency range.

What Do the Benchmarks Show?

Google says Gemini Decision 1 performed better across seven of the nine benchmarks it tested against Jev from TypeSafe AI and OpenAI’s Decisions API. The Google announcement dated November 11, 2026 says the model led across the benchmark suite and had the lowest calibration error, but the complete task list, baselines and scoring methodology are not published there in full.

AlphaLab’s examination of the published results gives more detail. On Mind2Web Multimodal, Google listed Gemini Decision 1 at 86.2% and Jev at 86.8%; on When2Call, it listed Gemini at 79.9% and Jev at 82.0%. The Top-1 ECE figure of 0.0250 applies to seven shared evaluations. ECE measures the difference between reported confidence and actual results, so that number does not mean each answer has a 2.5% error probability.

Google also says Gemini Decision 1 with the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the shared benchmark suite. That comparison applies to the shared suite, while Interesting Engineering reported on November 11, 2026 that the current OpenAI Decisions API supports up to 400 questions per request and starts at $0.10 per 1 million input tokens.

What Does the Gemini Decisions API Cost?

Google lists the Gemini Decisions API at $0.10 per 1 million input tokens with no output-token charges in its November 11, 2026 announcement. Google AI Studio provides a limited free tier, although the announcement does not state a free-tier quota.

Because shared input is billed once per request, several checks against the same material can share that charge, while each question adds its own instruction and criteria tokens. A 100,000-token paid request at the listed rate would cost $0.01 under the formula in AlphaLab’s pricing example, before any other service or processing charges.

What Does the Gemini Decisions API Mean for Business Workflows?

The Gemini Decisions API can support business workflows that end in a defined choice. A policy review can return yes or no, a routing request can select from fixed options, and a submitted item can receive a rubric score. Google lists tool selection, policy checks, model routing and context grading as intended uses.

For online visibility, BizScoreAI’s AI Scan and AI Audit show how an AI agent can read a business website and what it finds.

How Should a Business Calibrate and Monitor the Model?

A business can make each returned probability reviewable by recording the probability, confidence, input identifiers and final decision for representative cases. AlphaLab also recommends retaining a labelled sample, watching for shifts by language or media type, and checking results again when the error cost or data source changes.

  • Log the returned fields: Record the probability, confidence, selected option, input identifier and outcome for each reviewed decision.
  • Sample flagged cases: Set a fixed review rule for low-confidence answers and include a representative sample of accepted decisions.
  • Test the option set: Include a clear “other” option or a separate applicability check when none of the supplied choices fits.
  • Separate selection from authorization: Keep permission checks outside the model score and test adversarial inputs before a high-risk workflow uses the result.
  • Group the results: Review accuracy, confidence and human-review rates by language, media quality and uncommon category.

EmbeddingGemma 2 and Gemini Embedding 2 Releases

Two separate Google releases address the embedding layer used to compare and organize content. EmbeddingGemma 2 is an open-weight, multimodal embedding model that maps text, images, video frames and audio into one vector space. It has a 740-million-parameter footprint and can run with about 191 MB of active RAM for text-only weights or about 567 MB for the full multimodal model on a Pixel 11 Pro. Google says it supports zero-shot intent routing without training data or fine-tuning.

Glowing digital streams flow into a transparent sphere in a dark space

In a separate release, Google introduced Gemini Embedding 2 in public preview through the Gemini API and Vertex AI. It maps text, images, videos, audio and documents into one embedding space across more than 100 languages. Its stated limits include 8,192 input tokens for text, six images per request, video up to 120 seconds, and PDFs up to six pages.

EmbeddingGemma 2 and Gemini Embedding 2 organize text, images, video frames, audio and, for Gemini Embedding 2, documents within a unified embedding space. Gemini Decision 1 evaluates bounded choices and returns probabilities for those choices.

Key Considerations

  • The reported latency applies to contexts up to 16,000 tokens and requests containing one to 100 questions.
  • The shared input limit and the 1,000-question request limit do not establish the same maximum load for the reported latency.
  • The published table includes tasks where Jev scored higher.
  • A probability of 0.95 does not establish a 95% success rate without labelled evaluation data from the intended workflow.
  • The API is intended for closed-set questions and requires developers to define options, criteria and review thresholds.

What to Do Now

  1. List recurring business questions that require a yes-or-no result, a selected option or a defined rating.
  2. Inventory the records and inputs required for each question.
  3. Define acceptable answers, confidence thresholds and the human-review route with the process owner.
  4. Prepare a labelled set of representative inputs, including edge cases and cases with no suitable option.
  5. Test the workflow in the limited Google AI Studio tier and compare the results, latency and token charges with the current process.

FAQ

Can the Gemini Decisions API Replace a Human Reviewer?

Google documents that the confidence score can be used to route ambiguous cases to a fallback or human reviewer. It does not remove the need for a defined review path, and Google’s benchmark results do not establish reliability for a particular business workflow.

Does Using the Gemini Decisions API Affect Google Search or AI Mode Visibility?

No direct effect is stated. The Decisions API is a new endpoint for fast, closed-set decisions powered by Gemini Decision 1.

How Many Questions Can One Gemini Decisions API Request Include?

One request can include up to 1,000 independent questions against a shared multimodal input containing up to 256,000 tokens.

What Does the Gemini Decisions API Cost?

Google listed $0.10 per 1 million input tokens with no output-token charges on November 11, 2026. Google AI Studio also provides a limited free tier, although the announcement does not state a free-tier quota.

Which Question Types Does Gemini Decision 1 Support?

Gemini Decision 1 supports yes-or-no conditions, multiple-choice questions with 2 to 255 options, and ordered-rubric scoring.

SEOScanPro

SEOScanPro, which includes the AI visibility report

SEOScanPro runs a full technical audit of a site and shows the measured result behind every check. Open the AI visibility report.

Sources


See our editorial disclaimer for how our articles are produced.

🤖
Is your business visible to AI assistants?

Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.

Check Your Score →