News
FRONTIER NEWS / WHAT IS CHANGING NOW

Google Introduces Gemini Decisions API as Closed-Set AI Becomes Enterprise Infrastructure

Google’s Gemini Decisions API evaluates multiple closed-set questions against one shared multimodal input, signaling a broader shift from general-purpose generation toward specialized decision infrastructure governed by workflow-level accuracy, confidence thresholds and escalation controls.

FRONTIER RESEARCH / PORTABLE EDITIONS

Take the argument with you.

Read the publication-ready report. Generated from the canonical News article.

Frontier editorial art for Google Introduces Gemini Decisions API as Closed-Set AI Becomes Enterprise Infrastructure

What changed

On November 11, 2026, Google introduced the Gemini Decisions API, a new endpoint powered by its Gemini Decision 1 model and made available through the Gemini API in Google AI Studio.1 Independent reports corroborated the launch, model name, closed-set format and AI Studio availability.23

The API is designed to answer between one and 1,000 independent questions about shared input within a 262,144-token context window that includes both the input and the questions. It accepts text, images, audio, video and PDF files, while questions can use yes-or-no answers, fixed choices or ordered rubrics.4 This is materially different from asking a general-purpose model to produce an open-ended response: the endpoint constrains the output space and evaluates multiple specified criteria against the same underlying record.

Google lists pricing at $0.10 per million input tokens with no output-token charge. The shared input is counted once in a request, while each question contributes its own instructions and criteria.14 That creates a credible economic advantage when an enterprise currently resends the same document to multiple independent checks. Under the candidate report’s illustrative assumptions—a 4,000-token record and twenty 50-token checks—a batched request processes 5,000 tokens, costing $0.50 per 1,000 records. Twenty separate calls would process 81,000 tokens per record across the calls and cost $8.10 per 1,000 records, a 16.2-times ratio.5

That ratio is analysis of a particular request design, not a general Google benchmark. It depends on every separate call resending the complete 4,000-token record; savings will narrow when applications already cache, preprocess or avoid duplicating context.4 Google separately claims 150–250 millisecond end-to-end latency for contexts up to 16,000 tokens and says Gemini Decision 1 is nearly five times cheaper than Gemini 3.5 Flash-Lite across its shared benchmark suite.1 Those remain vendor-reported results and require validation under each organization’s context lengths, question counts and traffic patterns.

Notably, the headline also combines Google’s launch with a 63% document-pass result that does not belong to Gemini Decision 1. Tech Times attributes that result to Nace AI’s Drex Max v1.1 and says the underlying benchmark was published on November 8, three days before Google’s announcement.5 The same report relays a 33% Gemini result through another outlet but says it still requires confirmation. Neither figure provides a verified, like-for-like basis for judging Google’s new endpoint.

Why it matters

The CIO decision is no longer simply whether one model is cheaper or more accurate than another. Closed-set decision endpoints change the unit of evaluation from an individual prompt and response to an entire governed workflow: one shared record, a defined collection of questions, a document-level correctness requirement and a policy for handling uncertainty.

The first change is economic. Conventional cost comparisons often multiply a model’s token price by call volume. A shared-context endpoint makes application topology equally important. A workflow that repeatedly sends the same invoice, claim, contract or media file to separate checks may be paying primarily for duplicated input processing. Consolidating those checks can reduce that duplication, but the relevant metric is the proportion of total workload tokens attributable to repeated context—not the advertised price of an isolated request. The 16.2-times example demonstrates the mechanism, not an assured enterprise saving.54

The second change is how accuracy must be measured. If a workflow asks twenty questions about one document, per-question accuracy can conceal operational failure. A record may be unusable when even one required policy check is wrong. CIOs therefore need an all-answers-correct measure at the document or case level, segmented by document type, risk class, context length and number of questions. The unrelated 63% result cited in the candidate coverage illustrates why model-level headlines are especially weak procurement evidence: it concerns another vendor’s model, and its publication preceded Google’s launch.5

The third change is the role of confidence. Google’s documentation advises customers to calibrate confidence thresholds against labeled examples from their own traffic, adversarially test untrusted inputs and retain human or secondary checks for high-stakes actions.4 These controls make the escalation rate part of the business case. A model can be inexpensive per request but costly in operation if low-confidence cases overwhelm reviewers. Conversely, an endpoint that is not accurate enough for autonomous approval may still be valuable as a triage layer when it reliably separates routine cases from those requiring investigation.

The fourth change is architectural. The endpoint supports up to 1,000 questions and a large aggregate context window, but those limits should not be treated as a target design.4 Longer records and larger question sets can alter latency, cost, failure patterns and the operational consequences of a partially incorrect result. CIOs should evaluate the actual combinations their systems will submit rather than extrapolating from short-context vendor claims. They must also determine whether questions are genuinely independent. Where one decision depends on another, a single large batch may obscure sequencing, exception handling or policy logic.

Procurement consequently shifts from selecting a model to approving a controlled decision service. The approval package must connect shared-context savings to workload-level reliability, calibrated thresholds, review capacity and downstream consequences. That is a more demanding standard than a prompt demonstration, but it also gives CIOs a clearer path to measurable production value.

Frontier take

Closed-set AI should be treated as probabilistic decision infrastructure, not as a discounted version of a general-purpose language model. Google’s launch matters because it packages shared-context processing, constrained outputs and confidence-bearing decisions into a distinct endpoint. The strategic value is not that every classification task should move to Gemini Decision 1; it is that recurring enterprise checks can now be redesigned around a purpose-built execution pattern.14

That pattern should pass three gates. The first is topology: does the current workflow repeatedly process the same source material across independent checks? If not, shared-context pricing may offer little structural benefit. If it does, CIOs can calculate the avoidable duplicated input before conducting a model trial. This prevents an assumption-dependent ratio such as 16.2 times from being mistaken for a forecast.5

The second gate is workflow reliability. Evaluation must reproduce production context sizes, media types, question counts and answer constraints. It must score complete records as well as individual questions, because aggregate question accuracy does not reveal how often a document clears every required check. Vendor benchmark comparisons are secondary to a labeled internal test set representing the organization’s own traffic. Google’s own documentation reinforces the need for customer-specific calibration rather than universal confidence cutoffs.4

The third gate is control design. Confidence thresholds, adversarial testing and human or secondary review are not post-deployment refinements; they determine whether the service is economically and operationally viable.4 A production decision should specify which outcomes can proceed automatically, which require another machine check and which must reach a person. It should also quantify the review volume generated by the selected thresholds.

The resulting architecture is neither fully autonomous nor merely assistive. It is a governed layer that can batch routine judgments while routing uncertainty according to consequence. CIOs that evaluate this category through token prices alone risk moving cost from inference into exception handling. Those that jointly model duplicated context, whole-record correctness and escalation capacity can identify the narrower set of workflows where specialized decision endpoints provide a durable advantage.

Three moves for CIOs

  1. — Measure context duplication before selecting a decision model Map each recurring classification or policy-checking workflow as a context-to-question graph. For every record type, calculate how many times the same input is transmitted, what percentage of processed input tokens is duplicated and whether the checks are independent enough to share one request. Build the business case from the organization’s observed record lengths and question instructions rather than applying the published 16.2-times illustration as a forecast.54

    • Decision trigger: Start a batched-endpoint proof of value when the same source record is being resent to at least five independent checks and duplicated context represents more than half of the workflow’s input-token volume. Do not consolidate when checks require sequential dependencies that the application must explicitly enforce.
    • Why now: Google’s shared-input pricing makes duplicated context an addressable architecture cost, but the size of the opportunity depends on request design.1 Measuring duplication first gives the CIO a vendor-neutral baseline and reveals whether the workflow has the structural conditions needed to benefit from this category.
  2. — Replace per-answer benchmarks with a whole-record release test Create a labeled evaluation set drawn from actual production traffic and score both question-level performance and the percentage of records for which every required answer is correct. Segment results by input format, context-length band, question count and business-risk class. Include the current process as the baseline, then test the endpoint at the exact batch sizes proposed for production. Treat third-party headline results as contextual evidence only, particularly when they refer to a different vendor or predate the product being assessed.5

    • Decision trigger: Approve production use only when the all-answers-correct rate meets or exceeds the organization’s current process for every in-scope risk segment, and when no tested context-length or question-count band falls below the preapproved release threshold. If either condition fails, narrow the use case or retain the endpoint solely for triage.
    • Why now: Closed-set endpoints can return many individually plausible answers while still failing a document that requires every check to pass. The candidate coverage’s unrelated 63% Nace AI result demonstrates why CIOs need workload-specific, model-specific evidence rather than blended benchmark headlines.5
  3. — Budget escalation capacity as part of inference economics Calibrate confidence thresholds using labeled examples from the organization’s own traffic, then calculate the human-review or secondary-model volume produced at each threshold. Define separate routing policies for routine, consequential and high-stakes decisions. Adversarially test untrusted inputs, and require a human or second check before any high-stakes action identified by the organization’s governance process.4

    • Decision trigger: Block autonomous deployment whenever the calibrated threshold would exceed available review capacity, whenever high-stakes cases lack a mandatory second check, or whenever adversarial testing produces an unresolved path to an incorrect downstream action. Recalibrate when input mix, document templates, question wording or review capacity changes materially.
    • Why now: A low request price can be offset by an unmanageable exception queue. Google explicitly recommends customer-specific threshold calibration, adversarial testing and additional review for high-stakes actions.4 Integrating those controls into the original business case ensures that shared-context savings survive real operating conditions.

Sources