Google’s Gemini Decision 1 model scores up to 1,000 questions in one API call

Google DeepMind released the Gemini Decisions API Wednesday for developers who need software to make a series of small judgments quickly. A claims intake system, for example, could send a scanned form, photos of the damage and notes from a customer call in one request. It could ask which queue should get the claim, whether evidence is missing and how urgently someone should review it. The API returns answers and confidence scores that the application can use to move the case along.

Google’s new model, Gemini Decision 1, can answer up to 1,000 questions about the same material in one pass. Google charges $0.10 per million input tokens, with free output. That puts Google into a market where TypeSafe’s Jev already handles inexpensive classification, OpenAI’s Decisions API runs on GPT-6 Luna, and Amazon’s Strands Decider gives developers the option to run the model on their own infrastructure.

The enterprise case for Gemini starts with what goes into the request. A team handling recordings, footage and PDFs can evaluate those files together. A team routing short text tickets has cheaper hosted options, and the Artificial Analysis Decision Index puts Jev and Gemini close on overall quality.

One request can cover an entire checklist

The Decisions API guide describes three kinds of questions. A yes-or-no check returns the probability that a condition holds.

A multiple-choice question picks from two to 255 named options and reports the probability of each. A rubric question returns a weighted score and probabilities across a scale the developer defines.

The application can use those probabilities to route ambiguous cases to a person or ask a second model for a judgment.

The model scores the available options in one pass through the input.

It can read text, images, audio, video and PDFs together, and developers can adjust image and video resolution to control how much detail it sees and how many tokens the media consumes. The 262,144-token limit covers the shared evidence plus the instructions and choices for all questions.

Each question is evaluated independently. If a second decision depends on the first, the application can handle that dependency in code or make a follow-up call. Tasks such as drafting an explanation or pulling an arbitrary invoice number from a file still call for a separate generation or extraction step.

Google says Gemini Notebook uses the model to route queries to different models and levels of reasoning effort. The company reports up to 40% lower latency for routine queries in that workflow, where the routing step can reserve the more time-consuming models for harder requests.

Google’s pricing page counts the shared input once per request and bills the tokens each question adds. Take a 4,000-token record and 20 checks that add 50 tokens apiece.

At Google’s rate, one 5,000-token request costs about $0.50 per 1,000 records. Twenty separate calls that each resend the record would cost $8.10 per 1,000 records, before any retries or follow-up calls. The example assumes those token counts are the billed totals. Google charges for the full input again when it is reused in a separate request.

Developers can start with the Gemini Developer API’s v1beta endpoint and a no-cost tier in Google AI Studio. Google lists free-tier limits of 200 requests and 50,000 input tokens per minute. OpenRouter and Vercel AI Gateway also list the model. Vercel directs developers using audio, video or PDFs to its AI SDK 7 integration. Google says support for Gemini in Google Cloud is coming.

For teams testing with customer records, the Gemini API terms distinguish between paid and unpaid access. Google says it excludes paid-service prompts and responses from product improvement and retains logs for a limited period for policy enforcement and required disclosures. Its unpaid-service terms allow product improvement uses, with regional exceptions.

Google’s benchmark lead depends on the task

In Google’s launch benchmarks, Gemini Decision 1 led five of the seven tests that all four models ran. It scored 92.2% on JevBench, ahead of GPT-6 Luna at 90.9% and Jev 1.14 at 90.0%.

Jev led on the browser-task benchmark Mind2Web Multimodal, 86.8% to 86.2%, and on When2Call, which tests tool-use decisions, 82.0% to 79.9%.

On the two media tests, Google compared Decision 1 with Gemini 3.5 Flash-Lite. Decision 1 scored 71.8% to 67.7% on audio benchmark MMAU and 69.1% to 63.3% on Video-MME.

Decision 1 had the lowest calibration error in Google’s comparison, 0.0250, roughly half Jev’s 0.0494. Calibration measures how closely the model’s confidence tracks how often it is right.

It matters when a confidence threshold determines how many cases reach a reviewer. Google also reports that its shared tests cost $0.88 with Decision 1, $0.68 with Jev and $4.14 with Gemini 3.5 Flash-Lite through the Interactions API. Flash-Lite cost about 4.7 times as much as Decision 1 on that workload.

Google’s reported end-to-end latency of 150–250 milliseconds covers inputs up to 16K tokens, with little change as requests went from one to 100 questions. Google puts median model compute at 50–100 ms under those conditions. Teams planning to use the full context window or the 1,000-question limit should measure how long their own files and checklists take, including time spent uploading media.

The independent Artificial Analysis Decision Index v2.1 puts Gemini ninth of 18 at launch. It combines six evaluations, including document review, tool calling, knowledge and reasoning. Its latency figure is a weighted average of per-test medians, and the cost column estimates what a million tasks from that test mix would cost at the models’ listed rates.

Gemini led the independent suite’s knowledge and reasoning tests. On AA-Paperclip, which asks 15 questions about each résumé, research paper, invoice or receipt, Gemini got every answer right on 33% of documents, versus 35.5% for Jev and 63% for Drex Max v1.1.

A workflow that requires every field or check to be correct will care about that all-or-nothing measure even when individual answers score well. On Artificial Analysis’s When2Call run, Drex Max scored 94.6% accuracy, followed among these three models by Jev at 80.8% and Gemini at 75.2%.

Jev is cheaper for text routing, and OpenAI can reuse context across calls

TypeSafe prices Jev 1.14 at $0.042 per million input tokens with free output. Like Gemini, it can evaluate several questions about a shared record in one call. Jev’s standard limits are 64K tokens across the request and 32K across the shared material plus any individual question.

An experimental setting raises both limits to 262K. Jev can take text and structured JSON, and its opt-in image support lets developers attach a PNG or JPEG to the shared material or to an individual question. Version 1.14 also adds an experimental bounding-box answer for locating an object in an image. For short support tickets or routine tool selection, Jev’s price and its 92 ms result in the independent suite make it a good baseline.

OpenAI made its Decisions API generally available on Oct. 22. GPT-6 Luna is the sole supported model on the endpoint. A request can ask up to 400 questions and use text, images, native JSON, or records of previous tool calls as evidence. Images can also go into individual questions.

A developer comparing product photos against a shared reference can attach a different image to each question. For voice applications, OpenAI’s GPT-Live integration can trigger decisions from partial transcripts.

OpenAI charges $0.10 per million uncached input tokens and $0 for output, matching Google’s base rate. Its cache reads cost $0.01 per million tokens, with charges for cache writes. That gives teams an option when they repeatedly send the same policy manual, product catalog or reference image across separate requests. The reusable prefix needs at least 1,024 matching tokens, and content specific to a question sits outside that prefix. Regional processing and long-context pricing can also affect the bill. OpenAI documents Zero Data Retention for eligible customers and regional processing in the U.S., the European Economic Area and Switzerland.

Two other contenders are worth including where their inputs fit. Nace AI’s Drex Max v1.1 charges $0.03 per million input tokens and takes text, JSON and images, making the independent document and tool-use results worth checking against a company’s own work.

For audio and video, Cloudflare’s Clef-Omni takes both formats, offers a 64,000-token window on Workers AI and costs $0.15 per million input tokens there. Cloudflare also publishes the model weights, giving teams a second option if they need to run media evaluation in their own environment.

Amazon’s Strands makes self-hosting a real option

Amazon released Strands Decider 2B in October, as VentureBeat reported at launch. Its current Qwen3.5-based 2B checkpoint supports yes-or-no, multiple-choice and rubric questions over text.

It is available under Apache 2.0, with serving code and a training recipe. Teams can run it on a CPU or GPU and pay for their own compute and operations. That makes it relevant to companies that want the decision step to stay inside their own network.

Amazon reports 77.9% accuracy for the 2B checkpoint on the 231 public tasks in JevBench v1 using a 4,096-token window. Its earlier v19 checkpoint took 115 ms at the median and 299 ms at the 95th percentile for a single question on a local RTX 3090. Amazon’s later Gemma 4 variants range from 77.9% to 89.6% accuracy on that same public test set.

Google and Amazon report JevBench results from separate runs. A buyer weighing these options should test the same questions on both deployments and measure the full time from request to answer.

An Amazon internal ticket-routing deployment offers a better look at how self-hosting can pay off. Engineers ran Strands Decider E4B alongside Claude Sonnet 4.6 before switching which model applied the labels. They shortened the criteria for the 20 labels and trimmed the oldest comments on long tickets to fit the model.

Amazon reports average time per call fell from about 4.2 seconds to 0.9 seconds. At published AWS prices and its observed instance utilization, the team put the cost at about $500 per million tickets, down from $22,200 for the Sonnet workflow. A buyer considering the same approach would need to price its own capacity and the staff time to run it.

Where to put a decision model in production

Gemini is a good first pilot when one item arrives in several formats and triggers a checklist. A contact center could send a recording with order metadata and ask the model to identify the issue, choose a review queue and flag whether a required disclosure appears in the call.

In a photo-based damage workflow, the same request can check evidence from a form against several photos. In both cases, review the decisions individually. A strong queue-routing result says little about how often the model will miss a rare disclosure.

For developers building an agent, a decision model can sit just before a tool call. Ask whether the proposed tool is relevant, whether a required detail appears in the conversation and whether a person should review the action. Define an explicit escalation option for cases outside the available choices.

Keep access controls, approval requirements and transaction limits in application code. In one prompt-injection test reported by VentureBeat, a single planted field dropped Jev’s probability of blocking an action from 0.76 to 0.48.

Run a candidate in parallel with the existing workflow before letting its output change a customer record. Use labeled examples from real traffic, including rare cases and poorly captured media. Set confidence thresholds for each decision and each model.

On a multiple-choice question, Google uses the selected option’s probability as its confidence score. In TypeSafe’s example, Jev assigns a 0.61 probability to an option and returns a confidence score of 0.42.

Track the share of cases that go to review, errors among cases the model handles automatically, and end-to-end latency at the 95th percentile. For short text, run Jev or Drex on the same set.

For recordings and video, compare Gemini and Clef-Omni on the same clips and measure how many tokens each service bills. Include Luna when the application can reuse a long reference across requests and Strands when you can serve it inside your own environment.

Compare the cost and time per correctly handled case, including the cases that still require a person.