- 6 min read
- 8 key claims
- 2 sources cited
- For Developers, Businesses, AI makers
Overview
The Gemini Decisions API is a new endpoint in the Gemini API powered by Gemini Decision 1 (gemini-decision-1). It is designed for bounded, closed-set judgments such as tool selection, policy checks, model routing, and content grading, rather than open-ended text generation. A developer supplies one shared input and a set of questions; the endpoint returns probabilities, selected options or rubric scores, and a confidence score for each answer [1].
Three question types are documented. The noul type estimates the probability that a yes/no condition holds, choice selects among 2 to 255 options with normalized probabilities for each candidate, and score grades against an ordered rubric and returns a continuous weighted score with per-level probabilities. Simon Willison independently describes the same three question types and the noul terminology [1][2].
Google says the endpoint accepts text, images, audio, video, and PDFs up to 256K tokens, with tunable resolution on image and video inputs to manage visual detail and token cost. The input field uses the same Content block format as the Interactions API. The example request uses API version v1beta; Google says API keys are available through Google AI Studio, with Python, JavaScript, and REST guides [1].
Why it matters
The notable change is the unit of work. Instead of prompting a generative model and parsing prose, a developer defines questions, options, and rubric levels once and receives structured probabilities in a single call. Google says up to 1,000 independent questions can share one input, which shifts some repeated classification and policy-check logic from separate model calls or dedicated classifiers into one request [1].
The returned confidence scores are the mechanism for automation: Google describes routing ambiguous cases to a fallback or a human reviewer based on those scores [1]. Whether that pattern works in practice depends on how meaningful and well calibrated the scores are, which the evidence does not establish independently.
Opportunities
Developers could use a shared-content evaluation layer that asks several independent questions over one document, such as checking a policy condition, selecting a cleanup action, and assigning severity in a single request. The three question types support more than binary classification: choice covers 2 to 255 options, and score produces an ordered, weighted grade [1].
A confidence-gated routing pattern is also plausible: handle clear cases automatically and send lower-confidence cases to another model or a person. Google lists tool selection, policy checks, model routing, and context grading as intended use cases [1].
Google reports three internal uses. Gemini Notebook uses the API as a real-time router and reports up to 40% lower latency for everyday queries, though the baseline is not specified. A rapid canvas experiment in Stitch uses it to generate playable wireframes and apply voice or tap edits. CC from Google Labs uses it to select among more than 200 emoji reactions, which Google reports happens in a fraction of a second [1].
Risks
Reliability and calibration remain the main adoption concerns. Google says Gemini Decision 1 leads its benchmark suite for agent routing, reasoning, ranking, and multimodal understanding and has the lowest calibration error, but it does not name the benchmarks, baselines, calibration metric, or results, and no independent third-party test report is included [1]. The evidence also does not explain how confidence scores are derived or what thresholds should trigger human review [1].
Latency figures are vendor-reported without stated test conditions, regions, hardware, or network conditions [1]. Google reports 150–250 ms end-to-end latency and 50–100 ms median model-compute latency across multimodal inputs and contexts up to 16K tokens, and says latency stays nearly flat from 1 to 100 questions per request. The full advertised range extends to 1,000 questions, so scaling behavior above 100 questions is not demonstrated [1]. The 16K-token latency context and the 256K-token input limit cover different scopes, leaving latency beyond 16K tokens unknown [1].
Cost comparisons are incomplete. Google says the Decisions API is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API across a shared benchmark suite, but the benchmark composition is not identified and the comparison does not establish total cost for equivalent outputs [1]. The stated $0.10 per million input-token rate also does not include fallback models, human review, or surrounding infrastructure [1].
Applications may become dependent on Google's question schemas, confidence semantics, and Content block format, even though the input format is shared with the Interactions API [1]. A missed policy condition or a misrouted automated decision could carry operational or regulatory consequences, and no reliability figures are reported for individual modalities such as audio, video, or PDFs [1].
Context
Decision-model APIs differ from general chat or generation APIs: the developer defines the allowable answers in advance, and the service returns a bounded judgment. Willison places Gemini Decision 1 in a category alongside Jev and an OpenAI model, noting that the API resembles Jev in question types while extending coverage to images and audio [2].
At the document level, Google states a price of $0.10 per 1M input tokens with zero output tokens billed. Willison reports the same 10 cents per million input-token rate with free output [1][2]. The evidence does not state service limits, regional availability, or whether the beta label applies to the endpoint or only to the SDK API version.
Known, claimed, inferred
The documented interface is clear: three question types, a shared multimodal input, probabilities and confidence scores on every answer, and a Content block input format matching the Interactions API. The noul label is unusual, but Willison confirms it is intentional terminology carried over from Jev [1][2].
Google describes the service as scoring candidate options in a single forward pass rather than generating answer text token by token, which is the stated basis for its latency and zero-output-token billing claims [1]. That mechanism is also what makes batching many questions plausible in a single request, but the evidence provides no results comparing per-question quality in a batch against evaluating the same questions separately.
A reasoned inference is that the API's practical value depends less on raw generation speed than on the quality of its probability outputs. If confidence scores separate clear cases from ambiguous ones reliably, a single call can replace several classification steps and enable confidence-gated routing. If calibration is weak, the same design can misroute decisions. The current evidence supports the interface and the vendor's performance claims, but not that calibration claim.
The internal Google examples show that the service can be used as a control layer around other models rather than as an end-user generator. All three are self-reported, and the Notebook latency comparison does not name its baseline [1].
Limitations & uncertainties
There is no independent performance, calibration, latency, or cost test report. The evidence does not define calibration operationally, explain how confidence scores are derived, or specify reliability figures for audio, video, or PDFs [1].
Google's benchmark names, methodology, sample sizes, baseline configurations, accuracy, and underlying results are not provided, and the 'nearly 5x cheaper' comparison has an unspecified benchmark composition and does not establish total cost for equivalent outputs [1]. The latency figures lack test conditions, and the nearly flat latency claim covers 1 to 100 questions rather than the full 1,000-question maximum [1]. Latency beyond 16K tokens is unknown, and no results show whether batching preserves per-question quality [1].
Service limits, regional availability, and version-stability plans are not stated, and the evidence does not clarify whether the beta label applies to the endpoint or only to the SDK API version [1]. Thresholds for routing ambiguous cases to a human reviewer are unspecified [1]. Willison's excerpt ends with text, images, and audio, so his coverage of video and PDFs is incomplete rather than contradictory [2]. Evidence 112 is a truncated duplicate of the official announcement and adds no independently verified information [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementThe Gemini Decisions API evaluates up to 1,000 independent yes/no, multiple-choice, and rubric-scoring questions against one shared input in a single request.[1]
- Primary-source or vendor statementThe endpoint supports text, images, audio, video, and PDFs, with input context stated as up to 256K tokens.[1]
- Primary-source or vendor statementThe documented question types are noul, choice with 2 to 255 options, and score, which returns a continuous weighted score and per-level probabilities.[1]
- Primary-source or vendor statementEvery answer includes a confidence score from 0 to 1.[1]
- Primary-source or vendor statementGoogle reports 150–250 ms end-to-end latency and 50–100 ms median model-compute latency across multimodal inputs and contexts up to 16K tokens.[1]
- Verified factGoogle states a price of $0.10 per 1M input tokens, with zero output tokens billed; Willison reports the same rate with free output.[1][2]
- Primary-source or vendor statementGoogle says Gemini Decision 1 leads its benchmark suite and has the lowest calibration error, but the benchmarks, baselines, and underlying results are not identified.[1]
- Primary-source or vendor statementGoogle says the Decisions API is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API, without specifying the benchmark composition.[1]
Sources
- 1Gemini Decisions API: Fast, calibrated multimodal decisions●blog.googleretrieved Nov 18, 2026
- 2Gemini Decisions APIsimonwillison.netretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.