🔍 Read the full analysis: Gemini Decisions API: A Fast Approach To Calibrated Multimodal Decisions on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google announced the Gemini Decisions API on Nov. 11, powered by Gemini Decision 1, for evaluating up to 1,000 yes/no, multiple-choice or rubric questions against one shared multimodal input. Google reports 150–250 ms end-to-end latency and input pricing of $0.10 per million tokens, but the source provides no independent benchmark results or detailed methodology.
Google announced the Gemini Decisions API on Nov. 11, introducing an endpoint powered by its Gemini Decision 1 model that can evaluate up to 1,000 defined questions against the same multimodal input. The service is aimed at developers building workflows that need several bounded decisions—such as routing a request or checking a policy condition—rather than a separate generated answer for each one; Google’s performance and cost figures remain company-reported claims.
The endpoint accepts text, images, audio, video and PDFs, with Google saying a shared input can contain up to 256,000 tokens. Developers can submit yes/no questions, choose among 2 to 255 options, or ask for an ordered rubric score. The API returns confidence values from 0 to 1; choice questions also receive normalized probabilities for each option, while rubric questions return probabilities by level and a weighted score.
Google reports 150–250 milliseconds of end-to-end latency, including 50–100 milliseconds for model computation. The company says latency stays nearly flat from one to 100 questions in a request, and describes performance testing with multimodal inputs and contexts up to 16,000 tokens. These are Google’s figures: the supplied announcement does not include independent testing, full benchmark results or a detailed account of how the measurements were made.
Google lists pricing at $0.10 per million input tokens and says output tokens are not charged. It says the shared input is billed once per request regardless of the number of questions, with each question adding a small cost. The API is accessible through the Gemini API in Google AI Studio, which Google says offers a no-cost tier for prototyping. Pricing and availability details may depend on the applicable product terms.
Gemini Decisions API Fast, calibrated multimodal decisions
Google’s API scores up to 1,000 defined questions against shared multimodal input. Its speed, pricing and calibration claims await independent evaluation.
How it works
Gemini Decision 1 scores defined answers in one pass, instead of generating a separate text response for each question.
Provide text, images, audio, video or PDFs once for the batch.
Ask yes/no, multiple-choice or ordered rubric questions.
Get answers with confidence values; choice questions also return probabilities.
Where it could help
Several checks on the same material may need fewer repeated model calls.
Choose a path
Route requests to a model tier or fallback when confidence is low.
Check policy
Score defined conditions across documents, images or recordings.
Grade against a rubric
Return ordered scores for context without generating a separate answer for each check.
What the announcement supports
These figures and deployments are reported by Google; the source gives no independent evaluation.
- Latency stays nearly flat from one to 100 questions.
- Gemini Notebook reports up to 40% lower latency for everyday queries.
- Shared input is billed once per request.
- Benchmark names, datasets and detailed results.
- How confidence calibration holds across tasks and customer data.
- Performance across modalities and production workloads.
Try it in context
Developers can prototype through the Gemini API in Google AI Studio’s no-cost tier.
Batching Decisions Could Cut Repeated Calls
The proposed change is an API pattern for applications that need several answers about the same material. A developer could submit a document and related images once, then ask separate questions about policy, classification or next steps. If the reported behavior holds in production, evaluating those questions together could reduce repeated input processing and make routing workflows faster or less expensive than issuing separate generative calls.
Google says its Gemini Notebook product uses the endpoint to route queries and reports up to 40% lower latency for everyday queries. It also describes deployments in Stitch, CC from Google Labs and a safety layer for MCP tools in Search. The announcement does not provide outside customer results or detailed measurements for those uses, so the examples show Google’s reported applications rather than independently verified outcomes.
Confidence values could give teams a signal for sending uncertain cases to a fallback system or human reviewer. But a confidence score is not proof that an answer is correct, fair or safe. Its practical usefulness depends on whether it is well calibrated for a particular task and dataset, and on how the system behaves when inputs are ambiguous or outside the expected range.
Products Worth Considering
LG UltraGear 32-inch Curved QHD Gaming Monitor (2560×1440, 180Hz, 1ms, 1000R, HDR10)
DJI Mini 4 Pro Fly More Combo with DJI RC 2
Jackery Explorer 300 Portable Power Station with 40W Solar Panel, 292Wh
As an affiliate, we earn on qualifying purchases.
A Scoring API, Not a Text Generator
Many generative model APIs produce responses by generating tokens in sequence. Google presents Gemini Decision 1 as a different approach: it scores candidate options in a single forward pass, instead of generating a separate text response for each question. That design is intended for cases where the actions or categories are already defined, rather than open-ended tasks that call for a written answer.
Google compares the Decisions API with Gemini 3.5 Flash-Lite through its Interactions API, saying the new setup was nearly five times cheaper across a shared benchmark suite. The source material does not name the benchmarks or provide scores, workload details or cost assumptions, making the comparison difficult to reproduce. The claim should be treated as a company comparison, not as an independently established result.
The announcement’s examples include checking multiple conditions against shared material, such as reviewing a claim document alongside damage photos or matching a call recording with account information. These illustrate possible uses; the source does not report measured outcomes for those scenarios.
“Rather than generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass.”
— Google, in its announcement
Benchmark and Calibration Evidence Is Missing
The source does not publish benchmark names, datasets, scores, confidence intervals or an independent assessment of the model’s calibration. It also does not explain how calibration was measured, or whether confidence values remain reliable across different tasks, modalities and customer data. Those gaps limit what developers can conclude from the reported latency, cost and confidence figures.
Google’s latency description covers contexts up to 16,000 tokens for its performance claims, while the API is described as accepting inputs up to 256,000 tokens. The announcement does not specify how latency changes at larger context sizes or across each supported modality. It also does not establish how the service performs under heavier production traffic or when requests approach the stated question limit of 1,000.
The cost comparison with Gemini 3.5 Flash-Lite lacks enough published detail to reproduce it. The source also does not clarify how ambiguous inputs are handled or provide external results for the cited product deployments. These are open questions, particularly for teams considering the API in safety checks or other consequential workflows.
Test the API Against Real Workloads
Developers can start prototyping through the no-cost tier in Google AI Studio and access the endpoint through the Gemini API, according to the announcement. For production decisions, teams will need to test response times and costs against their own inputs and existing classifiers or model calls, then check whether confidence values reliably identify cases that require review.
Useful next evidence would include documented benchmark methods, calibration results across tasks and modalities, and customer measurements at production scale. The source material does not give a date for further benchmark disclosures or other product milestones, so it remains unclear when those details may be available.
Key Questions
What does the Gemini Decisions API do?
It evaluates yes/no, multiple-choice and rubric questions against a shared input, which can include text, images, audio, video and PDFs. Google says one request can include up to 1,000 questions.
How fast does Google say the API is?
Google reports 150–250 milliseconds of end-to-end latency, with 50–100 milliseconds attributed to model computation. The source does not provide independent measurements or full benchmark methodology.
How is the API priced?
Google lists a price of $0.10 per million input tokens and says output tokens are free. It says shared input is billed once per request, with each question adding a small amount; developers should check current product terms for applicable pricing.
Do the confidence scores show that a decision is correct?
No. A score between 0 and 1 is an output signal, not proof of correctness or safety. The announcement does not publish enough calibration evidence to establish how reliable those scores are across different tasks or datasets.
Can developers try the API now?
The announcement says the endpoint is available through the Gemini API in Google AI Studio, which has a no-cost tier for prototyping. The source does not specify a date for additional benchmark disclosures.
Primary source: Google AI · via ThorstenMeyerAI.com
