Gemini Decisions API: Google's Model That Gives Scores Instead of Text

On November 11, Google presented Gemini Decisions API, a service that does not produce natural-language answers: it assigns scores and probabilities to a closed set of options specified by the application. The model behind it is called Gemini Decision 1, and each response also reports a confidence level from 0 to 1, useful for sending doubtful cases to a human reviewer.¹² The service enters a market where OpenAI has already been present since last month with its own Decisions API.¹²

How does a decision model differ from one that writes text?

Gemini Decision 1 evaluates the candidate options in a single computation pass, the so-called forward pass, and returns probabilities and confidence scores. It does not decode the answer one token at a time, as a generative model does.¹ According to Agenda Digitale, the software fixes the list of possible answers in advance, the model examines the available information and returns a judgment in terms of probability. What to do next is decided by the software, through rules, thresholds and checks. A score is not an authorization: spotting a possible anomaly in a file does not mean it can be rejected automatically.² According to Agenda Digitale, the service is designed for repetitive operations of AI agents and enterprise software, such as classifying a request, identifying a priority or choosing the next step. Google cites as uses choosing a tool, checking compliance with a policy and routing a request to the appropriate model.¹²

How do questions, scores and confidence work?

A single request carries one shared context, which can mix text, images, audio, video and PDFs up to 256,000 tokens, and can contain up to 1,000 independent questions. Questions come in three types: noul, a yes/no check; choice, a selection among 2 to 255 options with the probabilities of all candidates; and score, an evaluation on an ordered scale.¹² The questions do not see each other and run in parallel, so latency stays almost unchanged from one to 100 questions. According to Google, this holds for inputs up to 16,000 tokens, while the declared overall latency is 150-250 milliseconds.¹² Agenda Digitale gives the example of an insurance claim with a report, photos of the damage and a recording of a phone call. The system can establish whether any document is missing, which office should handle the case and how urgently it should be reviewed. The most uncertain cases go to an operator, and the threshold separating the two paths is set by the organization, not by the model.² Google says it uses the API in Gemini Notebook, where daily requests are served with response times up to 40% lower, and in Search as a security layer against adversarial content and behaviors linked to MCP tools. These are figures communicated by the company, not an independent comparison.¹²

How does it compare with OpenAI and other providers?

OpenAI’s Decisions API is based on GPT-6 Luna, entered public beta on October 6 and became generally available on October 22. The limit is 400 questions per request, compared with Google’s 1,000, and it offers probabilistic checks, multiple choices and scores on a scale.¹² OpenAI claims its service delivers decisions up to ten times faster than the Responses API. It accepts text, images and JSON data, supports context caching and offers regional processing in the United States and the European Union. Google’s service, at launch, supports neither prompt caching nor the Batch API.¹² Among the usage examples cited by OpenAI is Datadog, which uses the Decisions API to assess the risk of shell commands produced by agents before they are executed.² The field is not limited to the two big names. TypeSafe AI presented Jev on September 15, describing it as a “System One” model, that is, non-generative. Microsoft-Decision-1 is in public preview in the Foundry catalog and accepts text and JSON. Agenda Digitale also includes Cloudflare, Liquid AI and Perplexity in the comparison.²

What do the figures declared by Google show?

Google states that Gemini Decision 1 leads on benchmarks for agent routing, reasoning, ranking and multimodal understanding. In its table, the model is ahead in seven of nine tests, but it trails TypeSafe’s Jev 1.14 in two: Mind2Web Multimodal, where it scores 86.2% against Jev’s 86.8%, and When2Call, where it reaches 79.9% against 82.0%.¹ The top-1 calibration error, which measures how closely probabilities match the actual frequency of outcomes, is 0.0250 for Google’s model. The other values in the table are 0.0494 for Jev 1.14, 0.0721 for GPT-6 Luna with OpenAI Decisions and 0.2104 for Gemini 3.5 Flash-Lite. This is the figure that matters when setting the thresholds above which an action triggers automatically. All figures are declared by the company.¹

What changes?

The academic TypedBench benchmark, dedicated to models of this kind, warns that poorly calibrated probabilities, or ones sensitive to wording, can lead the software to unwanted actions, without a textual justification that operators can inspect. In tests on a hosted model, the probabilities were systematically underestimated and, with asymmetric costs, using them could produce worse outcomes than following the most likely answer. The study does not attribute these results to Google’s service.³

This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.

Sources

Footnotes

1. Gemini Decisions API: Google sfida OpenAI con un modello che non genera testo, Hardware Upgrade | il sito italiano sulla tecnologia, 2026-11-12 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

2. Modelli decisionali AI, Google sfida OpenAI: chi offre cosa, Agenda Digitale, 2026-11-12 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

3. [2610.11392] TypedBench: A Benchmark for Calibration, Framing Sensitivity, and Cost in System One Decision Models, arxiv.org ↩