Gemini Decisions API: What a Routing Request Costs
Google’s new Decisions API bills shared input once per request. See its limits and a worked example of ten checks over the same input.
Google introduced the Gemini Decisions API on November 11, 2026. It lets developers send one input with up to 1,000 independent yes/no, choice or scoring questions and receive probabilities without billed output tokens. The paid Standard rate for Gemini Decision 1 is $0.10 per million input tokens. The cost advantage for repeated checks comes from billing the shared input once in a combined request; each question still adds input tokens. Those details come from Google’s launch announcement, API documentation and pricing page.
What changed on November 11
The new endpoint evaluates bounded decisions over text, images, audio, video and PDFs. A developer can ask whether a support message requests a refund, which team should receive it, and whether it needs a policy review in one call. The questions share the message or document, but Google says they are evaluated independently: one question cannot use another question’s answer.
That distinction matters when designing a workflow. If the second decision depends on the first, the application must combine or sequence the results in code. The endpoint returns a yes/no probability, a choice among supplied options, or a score against a defined rubric. It does not write a customer reply or provide a generated explanation. Google’s comparison with structured output points to a text-generating API for extracted values, tool-call arguments and multi-step reasoning.
Developers building routers, content checks or repeated evaluations over the same record are the likely users. Existing generation calls need not change merely because this model exists. First identify a decision that has a small, explicit answer set and a known fallback when the model is uncertain.
The bill includes every question
Google’s pricing table for Gemini Decision 1 lists these current terms for the direct Gemini Developer API. The price applies across the supported input modalities. The free tier is listed, but availability and rate limits for a particular account were not measured here.
| Item | Published term |
|---|---|
| Paid Standard input | $0.10 per 1 million tokens |
| Billed output | 0 output tokens |
| Questions in one request | 1–1,000 |
| Choice options per question | 2–255 |
| Context across input and questions | 262,144 tokens |
| Context caching and Batch | Not supported |
The documentation’s limits section supplies the question, option and context limits. Google says the shared input is billed once per request, while each question contributes its own instructions, criteria and formatting tokens. A second request with the same shared input bills that input again. The API does not support context caching, Batch, Flex or Priority inference, according to its limitations section.
Zero output-token billing therefore does not mean a free paid-tier request. It means the input total drives this model’s token bill. If your application sends images or audio, inspect the API’s reported input-token usage by modality rather than assuming a file has the same token cost as its text description. The response reference describes a usage object with input counts and zero output tokens.
One input, ten checks: a worked example
Suppose a team evaluates 10,000 records. Each record has 2,000 billable shared-input tokens. Ten independent questions add 100 billable tokens each, including their instructions, criteria and formatting. These are illustrative counts, not a benchmark, API test or forecast.
| Request design | Tokens for 10,000 records | Cost at $0.10 / 1M |
|---|---|---|
| One request with ten questions | (2,000 + 10 × 100) × 10,000 = 30,000,000 | $3.00 |
| Ten requests, one question each | 10 × (2,000 + 100) × 10,000 = 210,000,000 | $21.00 |
In this example, grouping saves $18.00 in direct Decisions API input charges because the 2,000-token record is sent once instead of ten times. The calculation holds the question tokens and unit price constant. It does not compare the Decisions API with another model, include retries, or assume a Batch discount. Real token counts can differ from these round numbers, especially for media. Log the usage returned for representative inputs, then multiply observed billable tokens by the published rate. The site’s API cost calculator can help with a general token-budget worksheet, but its model list should be checked before using it for this new endpoint.
Where a decision call fits
A combined request makes sense when several independent checks use the same input: route a ticket, flag possible policy issues and grade relevance against separate criteria. It is a poor fit when the next question needs the previous answer, or when the product needs a written explanation or an extracted value. In those cases, code can sequence calls or hand the case to a generation model. The total workflow bill is decision calls + any escalated generation calls + retries; no escalation rate was measured here.
A choice result always distributes probability among the options you supplied. Google advises adding an “other” option or a separate scope check when none of your named choices may fit. Otherwise the highest-probability option can look usable even for an out-of-scope record. Put a threshold and a human or alternative-model path behind actions with meaningful consequences. Test those thresholds on labeled examples from your own workflow, including ambiguous and out-of-scope cases. Google describes its probabilities as calibrated, but that claim is not a measurement of your data.
Before replacing an existing routing step, count its current requests and token usage. Then compare a combined Decisions call on the same representative records. Check the answer quality, fallback frequency, latency and total bill in your own logs. Google reports latency figures in its announcement; this article does not claim an independent speed test. If you use a gateway, verify its bill separately: Vercel says it offers Gemini Decision 1 on AI Gateway, but Google’s direct API rate is not a quote for a reseller bill.
The DailyTech model tracker lists Gemini Decision 1. For broader context, see AI API prices per million tokens and what an AI evaluation run really costs. Those explain other pricing and evaluation choices; this launch’s special variable is how many independent questions can reuse one input within a request.
Sources
- Google’s November 11 announcement — launch, task and vendor performance claims.
- Google Decisions API documentation — request design, limits, response and limitations.
- Google Gemini Developer API pricing — direct API rate and billing rules.
- Vercel’s AI Gateway changelog — integration availability and supported input paths.
All pages read on November 14, 2026.
Free toolWhat will 10,000 chats a month cost?Estimate my bill
