Google’s new Gemini Decisions API, announced November 11 and open to developers through Google AI Studio, can score 20 checks on one 4,000-token record for about $0.50 per 1,000 records instead of $8.10, because it bills the shared input once per request, according to a worked example from Implicator that applies Google’s $0.10 per million input-token price. The catch for developers is reliability: in Artificial Analysis testing published November 8, the leading decision model, Nace AI’s Drex Max v1.1, passed 63 percent of documents on which all 15 required answers had to be right.

The model behind the endpoint, Gemini Decision 1 (model ID gemini-decision-1), does not generate text. Google DeepMind’s Min Choi and Felix Yu describe it as a tool for tool selection, policy checks, model routing and context grading, all closed-set decisions. One request can carry up to 1,000 independent questions over a shared input of up to 256K tokens that can mix text, images, audio, video and PDFs.

Every answer carries a confidence value between 0 and 1 so that code can send ambiguous cases to a fallback model or a human reviewer. In an example reported by The Decoder, a photo of a kitchen counter was checked with three questions at once and returned 0.95 confidence for a stain, 98 percent for the cleaning action and a severity score of 1.19 on a scale topping out at 2. A no-cost tier in Google AI Studio is available for prototyping.

Google arrives 57 days after TypeSafe AI’s Jev entered early access September 15. OpenAI’s Decisions API became generally available October 22 at the same $0.10 per million uncached input tokens that Google now charges. Teams weighing the three face two practical questions: how to structure requests so shared input is billed once, and how to confirm that a batch of answers is right often enough to trust.

Shared-Input Billing Turns 20 Checks Into One 5,000-Token Request

Google bills the shared input once per request, and each added question costs only its own instruction and criteria tokens. Output is free, and Google lists no thinking-token bills and no separate audio transcription or OCR fees.

Implicator’s worked example shows the effect. A 4,000-token record with 20 checks of about 50 tokens each travels as one 5,000-token request and costs about $0.50 per 1,000 records. Sending the same 20 checks as separate calls, each resending the record, costs about $8.10 per 1,000 records, or 16.2 times as much.

Billing structure differs by vendor, and the gap is large. Artificial Analysis found that GLiDE v1.2 and Decider Flash v1.2 bill input for every question, about 63,000 input tokens per AA-Paperclip task against about 9,000 for Drex Max v1.1, so GLiDE v1.2 costs nearly 8 times as much despite a list price about 33 percent higher. Perplexity’s smaller Flash model also bills input again for every question, according to a pricing roundup from Parallel. OpenAI’s endpoint, by contrast, shares one input across up to 400 questions per request. Developers can ask each vendor one question before signing: is input billed once per request or once per question?

Single-Pass Scoring Replaces Token-by-Token Decoding, Google Says

Google says Gemini Decision 1 scores candidate options in a single forward pass instead of writing an answer token by token. Google DeepMind says the model “scores options without token-by-token decoding,” which accounts for its speed, according to SiliconANGLE’s launch coverage. That also explains the pricing: with nothing generated, there are no output tokens to bill, so the shared input is the whole cost.

Google’s launch post does not detail the model’s architecture, so how Gemini Decision 1 reads and scores options is not public. Open-model experiments by developer Victor Dibia show the general mechanism class: a model reads the prompt once, scores each candidate option, then normalizes the scores so they sum to 1. Those tests used open Qwen models, not Gemini.

The same tests expose the engineering tradeoffs that a one-pass scorer carries. Cost grows with the number of options times the number of questions, accuracy fell to 64 percent at 77 options for a 7-billion-parameter instruct model, and a fitted temperature adjustment cut calibration error from 0.10 to 0.03. Google caps choice questions at 255 options, but whether Gemini Decision 1 shows the same pattern is untested in public.

Leading Decision Model Passes 63% of 15-Answer Documents in Independent Test

Artificial Analysis built AA-Paperclip from 1,500 private documents, including résumés, receipts, invoices and research papers. Each document carries 5 decisions and 10 fact extractions, and the primary metric is the share of documents on which all 15 answers are correct. Claude Haiku 5.5 (Max) passed 92 percent and GPT-6 Luna (Max) 86 percent. Among decision models, Drex Max v1.1 passed 63 percent, GLiDE v1.2 62 percent and Decider v1.2 55 percent. Drex Max v1.1 answered 97 percent of individual answers correctly, and decision models mostly lose documents on decisions rather than on fact extraction.

Gemini Decision 1 does not appear in the text of Artificial Analysis’s November 8 write-up, which was published three days before Google’s launch. Implicator reports that Gemini passed 33 percent of AA-Paperclip documents, against 35.5 percent for Jev and 63 percent for Drex Max, and ranked ninth of 18 on the index at launch. That is a single outlet’s account of independent data, and Artificial Analysis’s live leaderboard is the place to confirm it.

Simple arithmetic shows why the all-answers rule matters for batches. If 15 answers were independent and each were 97 percent likely to be right, a document would pass about 63 percent of the time, which happens to match Drex Max v1.1’s measured pass rate. Real errors correlate: 99 percent per-answer accuracy would predict about 86 percent under independence, yet Claude Haiku 5.5 (Max) measured 92 percent. The arithmetic is an illustration, not an Artificial Analysis result, and it carries one consequence for the 1,000-question batch: every added question is another way for a record to fail.

Confidence compounds the same way. Artificial Analysis computes a document’s confidence as the product of its 15 answer probabilities. By the same arithmetic, 15 answers that each carry 0.95 confidence multiply to about 46 percent, so a per-answer threshold that looks safe can hide a record that is more likely wrong than right.

Google’s own benchmark claims arrive without scores in the announcement text. Google says Gemini Decision 1 leads its suite on agent routing, reasoning, ranking and multimodal understanding with the lowest calibration error, and its blog shows the comparisons only as images. SiliconANGLE reports a lead over Jev and OpenAI on seven of nine benchmarks, while Implicator counts five leads in the seven tests that all four models ran.

Gemini Returns 0.92 Where Jev and OpenAI Return 1.0, Moving Threshold Cutoffs

Developer Simon Willison tested Gemini Decision 1 on a video of a stomping robot with four yes/no questions. It returned 0.9234 for a robot, 0.9230 for the USA, 0.0833 for France and 0.0498 for a dragon. Jev and OpenAI return exactly 1.0 or 0.0 on yes/no questions, while Gemini “insists on being a little less confident than that,” Willison wrote. His post is a single test run with no ground-truth labels, so it shows a difference in style, not in accuracy.

The style difference decides where a developer sets a cutoff. A threshold of 0.95 that works for a model that snaps to 1.0 would reject the 0.92 answers in Willison’s test, and a threshold of 0.90 would accept them. Google reports the lowest calibration error in its suite, and Implicator, relaying Google’s comparison, lists Gemini at 0.0250 against 0.0494 for Jev. Those figures are Google’s.

Independent work on a sibling model shows why such figures do not transfer. A UCLA preprint by Lingsen Meng audited Jev 1.13.0 on two public classification tasks and found that calibration error shifted by 0.071 between two ways of posing the same question. About 72 percent of answers on a four-class task sat at exactly 1.00, and because probabilities are rounded to 0.01, a threshold on the top probability offered only 53 operating points. The audit did not test Gemini Decision 1.

Artificial Analysis now rewards calibration directly. Its Decision Index v2, published November 8, scores each answer as 50 × (1 + c − (p − c)²), where p is the probability assigned to the chosen answer and c is 1 if the answer was right and 0 if wrong, which equals 50 × (1 + accuracy − Brier score) across an evaluation. The Brier score, proposed by Glenn W. Brier in 1950, measures the squared gap between forecast probability and outcome. In the index, Celeris-1.1 Decision passes 24 percent of AA-Paperclip documents against 36 percent for Clef yet ties it on Decision Score because its Brier score is 0.15 against 0.27. OpenAI’s own guide tells developers to set thresholds from labeled examples in their application.

Latency Claim Covers 16K-Token Contexts, Not 256K-Token Input Limit

Google puts end-to-end latency at 150 to 250 milliseconds, with median model compute of 50 to 100 milliseconds, and says latency stays nearly flat from 1 to 100 questions per request. The figure applies to multimodal inputs and contexts up to 16K tokens, not the full 256K-token input or the 1,000-question limit. Implicator advises teams that plan to use full-size inputs or 1,000-question batches to measure their own files and checklists.

Independent timing exists for rivals but not yet for Gemini. Artificial Analysis lists time per task of 92 milliseconds for Jev 1.14 and 275 milliseconds for Drex Max v1.1.

How Much Does Gemini Decisions API Cost Against Jev and OpenAI?

Google charges $0.10 per million input tokens and nothing for output, and Willison notes it is priced like GPT-6 Luna, the same list price OpenAI charges for uncached input. Jev lists at $0.042 per million input tokens, according to TypeSafe AI founder Diogo Almeida. Microsoft priced Microsoft-Decision-1 at $0.042 when it entered public preview in Microsoft Foundry on October 9.

Google’s own shared test run put Decision 1 at $0.88, Jev at $0.68 and Gemini 3.5 Flash-Lite through the Interactions API at $4.14, according to Implicator’s account of Google’s comparison. Google says the service is nearly five times cheaper than Gemini 3.5 Flash-Lite across that suite. Google’s argument for paying more than Jev rests on input type: Gemini Decision 1 accepts text, images, audio, video and PDFs, while Jev takes text and structured JSON with opt-in image support and OpenAI takes text and images.

The rivals shipped within weeks of each other. Cloudflare released Clef and Clef-flash under an Apache 2.0 license on October 1. TypeSafe announced a Series A of $870 million at a $7.5 billion valuation on October 9. OpenAI’s Decisions API entered public beta on October 6.

Google Search Uses It to Flag Adversarial MCP Tool Content; Notebook Cites 40% Lower Latency

Google says it runs the API internally in four places. Gemini Notebook uses it as a real-time router that sends queries to the right model tier and thinking level, with up to 40 percent lower latency for everyday queries. Google Search uses it as a safety layer that detects adversarial behavior and content in MCP tools. Stitch uses it for an experiment that generates playable wireframes in seconds, and CC from Google Labs uses it to pick an emoji reaction from more than 200. The latency figure is Google’s own claim.

The safety use deserves a caution. HiddenLayer researchers reported in November 2025 that appending a chosen token sequence, such as “=coffee,” could flip a text-classifier guardrail’s verdict from malicious to safe. That research targeted other guardrail classifiers, not Gemini Decision 1, and the finding that matters here is general: a classifier that gates tool content needs adversarial testing and a fallback path, which the confidence value is meant to supply.

Developers get an API key through Google AI Studio, with Python, JavaScript and REST guides in the Decisions API documentation. Integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, and Google lists Gemini in Google Cloud as coming soon, with a sign-up form.

Measure All-Answer Pass Rate on Own Records Before Trusting Batch Savings

The savings are real on paper: shared-input billing turns a 20-check job from $8.10 to $0.50 per 1,000 records, and Google’s $0.10 rate matches OpenAI’s. The risk is just as concrete: the leading decision model in Artificial Analysis’s document test passes 63 percent of 15-answer documents, and Gemini’s reported 33 percent is a single-outlet figure that still needs confirmation. A team should therefore run its own labeled records through the API, count how often every answer on a record is right, and set confidence cutoffs from that data instead of copying a threshold from another vendor. Teams should also test full-size inputs beyond 16K tokens, confirm how each vendor bills shared input, and keep a human or fallback model ready for records whose combined confidence falls short. Whether Google’s calibration holds up outside its own benchmarks is the question that independent leaderboards and developer tests will answer.

Frequently Asked Questions

How much does Gemini Decisions API cost?

Google lists $0.10 per million input tokens and no charge for output, bills shared input once per request, and offers a no-cost tier in AI Studio for prototyping. Jev lists at $0.042 per million input tokens and OpenAI at $0.10 for uncached input.

How is Gemini Decision 1 different from a Gemini language model?

A language model writes an answer token by token. Gemini Decision 1 scores options that the developer supplies and returns probabilities and a 0-to-1 confidence for each answer, across three question types: yes/no, choice from 2 to 255 options, and a score against an ordered rubric. Google’s claim that it is nearly five times cheaper than Gemini 3.5 Flash-Lite covers cost on its shared benchmark suite, not speed.

Does batching more questions on one record raise the chance of a wrong record?

It can. Artificial Analysis counts a document as passed only when all 15 of its answers are right, and no decision model it lists passes more than 63 percent of documents. Fifteen independent answers that are each 97 percent likely to be right would pass about 63 percent of the time. Open-model tests by Victor Dibia found that fitting a temperature adjustment cut calibration error from 0.10 to 0.03, which suggests labeled examples from your own task can improve confidence values.

Can Gemini Decision 1’s confidence score serve as a routing threshold without testing?

Not safely. Google says the value lets code route ambiguous cases to a fallback model or human, but Willison’s test shows Gemini returning values near 0.92 where Jev and OpenAI return 1.0. OpenAI’s guide likewise advises setting thresholds from labeled examples in your own application.