Gemini Decision 1 is designed for a different job from a chatbot: evaluate a defined question and return a result that software can use. Google's model card describes yes/no, choice and rubric questions over text and media. It produces no output text tokens. For teams classifying messages, routing requests or checking a document against several rules, that changes both the interface and the bill.
The attraction is a simpler path from input to a proposed action. The risk is confusing a convenient result with permission to act. A score can help decide which queue should receive a request; it should not silently become the authority to issue a refund, reject a customer or alter a shared record. The model and the application's decision policy remain separate responsibilities.
- 01Input determines the model chargeGoogle lists $0.10 per million input tokens, with shared input billed once within a request and no output token charge.
- 02Questions still consume tokensInstructions, criteria and formatting add to the bill and the shared context limit.
- 03The maximum is not the operating targetUp to 1,000 questions fit only if the complete request stays within the 262,144-token context window.
- 04Confidence needs local evaluationA provider's calibration description is not proof that a chosen automation threshold is safe for your workload.
01 — Job selectionUse a decision model for a bounded question
A decision model is most useful when the possible answers are known before the request is sent. A support message might belong to billing, delivery or general enquiries. A document might satisfy a clearly defined requirement. A rubric might score whether a response includes an answer, supporting evidence and a next step. These are proposed application examples, not measured performance claims.
The boundary matters because a classification cannot replace every part of a workflow. If a customer needs a written explanation, a separate response still has to be produced or selected. If the available categories omit the correct answer, a well-formed response can still send the request to the wrong place. Include an escalation path for ambiguous or unsupported cases instead of forcing every input into a convenient business label.
The earlier Microsoft and OpenAI decision-model comparison explains the broader product category. Google's entry should be evaluated against the same business outcome: correctly handled requests at an acceptable total cost. A lower model charge is useful only if the surrounding process remains accurate and supportable.
Choose a known category
Route an input among defined destinations, with an explicit fallback when none is appropriate. Evaluate mistakes by their operational consequence.
Explain or create
Keep a generative model or a human in the process when the task requires a reasoned explanation, a new document or an answer outside a predefined set.
02 — Rate cardCalculate the request before the question price
Google's pricing page, retrieved November 17, 2026, lists $0.10 per million input tokens across supported modalities. Shared input is billed once per request; each question adds its own instructions, criteria and formatting. Output usage is zero. The rate card also says context caching and the Batch API are unavailable for this model.
Here is an illustrative calculation, not an invoice. Assume a shared text input of 10,000 tokens and 100 questions adding 100 tokens each, including their request overhead. The complete request contains 20,000 billed tokens. At $0.10 per million, its model charge is $0.002. Dividing that amount by 100 yields $0.00002 per question. If the actual encoded questions are longer, the result changes; use returned usage rather than a character-count guess for an operational budget.
Sending those same 100 questions as 100 separate calls repeats the input. Under the same assumptions, that is 1,010,000 tokens and $0.101. The difference comes from request grouping, not a special per-question discount. Both calculations exclude application hosting, retrieval, storage, retries and human review. They also assume text token counts rather than converting a minute of audio or a PDF page into an invented token estimate.
| Illustrative request pattern | Input tokens | Model charge |
|---|---|---|
| 100 questions sharing one input | 20,000 | $0.002 |
| 100 separate requests | 1,010,000 | $0.101 |
| One million billed input tokens | 1,000,000 | $0.10 |
03 — Request limitsBudget for the complete set of questions
The Decisions API guide allows 1–1,000 questions per request within a 262,144-token window covering shared input and all questions. Choice questions accept 2–255 options; rubric questions accept 2–255 levels, with fewer recommended for most rubrics. A long document and a large question set compete for the same capacity.
Do not design around the headline question maximum alone. Imagine checking a long supplier document against a large policy library. The instructions may be longer than the document, especially if every question repeats definitions and examples. Count the request you will actually send, keep the question definitions versioned, and decide what happens when an input exceeds the accepted size. Quietly truncating the document can remove exactly the clause the check was supposed to find.
Separate independent checks from dependent reasoning. The API evaluates questions against shared input without letting them see one another's answers. If the second question depends on the first result, express that dependency in application code or a later request. Merely placing the questions in a numbered order does not create a reasoning workflow between them.
A request-count limit and a token limit constrain different things. Validate both, and retain the input version so a result can be traced to the material the model actually received.
04 — Decision policyChoose thresholds from the cost of mistakes
Google describes the returned probabilities as calibrated. That is a provider claim about the model, not evidence that a particular cutoff delivers a known error rate on your customer data. Calibration means that confidence should correspond to observed correctness over an appropriate set of cases; it is not a guarantee for every individual result. Keep the distinction explicit when presenting the system to a business owner.
Build a labelled evaluation set containing ordinary, ambiguous and costly cases. For each proposed threshold, inspect both errors that trigger an action incorrectly and errors that miss a required action. A route to a general support queue may tolerate a different trade-off from a route that suspends an account. Do not reuse one threshold simply because it looks conservative on a dashboard.
The API's yes/no probability range is 0.01–0.99. A rule requiring a value above 0.99 will never fire; a rule accepting 0.99 still needs evaluation. Keep an abstention path for cases that do not meet the acceptance criteria. The human-review cost guide explains why review demand belongs in the economic calculation alongside inference cost.
For every automatic action, record the question version, threshold, relevant error types and fallback. The business should be able to explain why the rule is acceptable without pointing only to the model's confidence label.
05 — Data handlingChoose the tier before using sensitive inputs
The pricing page distinguishes a free tier whose content is used to improve Google's products from a paid tier whose content is not used for that purpose, subject to the linked terms. That distinction should be reviewed before a prototype receives customer documents. It is not a claim that a paid request has no retention, no logging or every compliance property a business might require.
Start evaluation with material the team is authorised to use and an explicit account configuration. A developer's successful free-tier experiment does not establish the production organisation's data terms or capacity. Record the project, billing status and applicable policies in the deployment decision. These are operational checks a team can perform without exposing private customer content merely to see whether a demo works.
Multimodal support also changes what enters the system. A PDF may contain identifiers beyond the paragraph being classified; an image may include background information irrelevant to the question. Supply the information needed for the task and apply the organisation's existing handling rules. A low per-token price should not become a reason to send entire archives when a smaller, authorised input would answer the question.
Not used for product improvement is a specific data-use statement. It should not be expanded into an unsupported promise about retention, residency or legal suitability.
06 — Failure handlingKeep an error distinct from a negative answer
The API guide documents request-validation errors and safety-blocked requests. Neither is an ordinary classification result. If the application receives an error, it must not interpret an absent answer as false, safe or approved. Preserve that distinction in the data model and in the queue a person sees when a request needs attention.
A hypothetical document-routing service illustrates the problem. If a malformed question causes the request to fail, defaulting to the routine queue can conceal the failure from operators. A separate unresolved state makes the missing decision visible. Retrying may be appropriate for a temporary failure, but repeating an invalid request without changing it only increases delay and noise.
Test partial application failures as well as model results. The scoring call may succeed while the downstream record update fails. Store enough identity to avoid applying an action twice on retry, and keep the original decision with its input and policy version. These are application-design recommendations; the decision endpoint does not provide a complete transaction system for the business actions that follow.
Distinguish an accepted decision, a decision sent for review and a request that failed. Collapsing them into a single true/false field discards information the operator needs.
07 — AdoptionPilot one repeated decision before expanding
Choose a high-volume, bounded task with a visible ground truth and a reversible outcome. Run the proposed decisions alongside the existing process before allowing them to trigger actions. Review disagreements, especially where the model appears confident. The purpose is to learn whether the definition of the task is sound and whether the error trade-off is acceptable, not to produce a favourable demonstration.
Compare total operating cost with the current approach. Count the model requests, downstream work, review time and correction effort. The open decision-model comparison provides context for alternatives, but the winning model for a specific process still depends on its inputs and acceptance criteria. A cheaper call that creates more expensive corrections may be the wrong choice.
Digital Applied's AI transformation service can help connect a model evaluation to the workflow it will change. Gemini Decision 1 is worth testing where repeated questions share substantial input and do not need generated explanations. Move beyond that pilot only after the team can account for both the savings and the mistakes.
Expand when a defined task meets its accuracy, review and cost targets on representative inputs. A working API call is the start of the evaluation, not the result.
Price the workflow, then test the decision rule
Gemini Decision 1 makes shared-input classification inexpensive at the published token rate. Its practical value depends on clear questions, measured thresholds and reliable error handling. Start with one bounded decision and calculate the cost of getting it right, including the human work that remains.