Google Launches Gemini Decisions API, Judging Up to 1,000 Questions per Call with Free Output Tokens
Google has launched the Gemini Decisions API, invoked through the Gemini Decision 1 released the same day, which can make yes/no judgments, multiple-choice selections, and graded scoring on up to 1,000 preset questions at once, supports text, image, audio, video, and PDF inputs, and does not charge for output tokens. Pricing is $0.1 per million input tokens; Google says that with 16K-token multimodal input, median compute latency is 50 to 100 milliseconds and end-to-end latency is 150 to 250 milliseconds, at about one-fifth the cost of Gemini 3.5 Flash-Lite. Developer Lesia Kaplyuk tested 54 calls; all 30 customer service tickets and 8 multilingual tasks were classified correctly, and processing 100 questions in a single call took 0.41 seconds, but when no fallback option was set, the model forced out-of-scope inputs into existing categories. The API is already used in Google internal products such as Gemini Notebook and Stitch, and Google Search also uses it for security protection of MCP tools.
Why read this
AI score 73
The original article provides pricing, latency, and developer test data, which can be used to assess the applicable boundaries of this type of judgment API in classification and routing scenarios.
Topics
Body · AI translation
Smart Things news on November 12: Today, Google announced the launch of the Gemini Decisions API, which can be called through the newly released model Gemini Decision 1 on the same day. Unlike traditional generative models, Gemini Decision 1 can directly evaluate input content and make judgments on up to 1,000 preset questions at once. The model supports multimodal inputs such as text, images, audio, video, and PDFs, and does not charge for output tokens.
In terms of performance, Google says that in multimodal input scenarios of up to 16K tokens, Gemini Decision 1 has a median model compute latency of 50 to 100 milliseconds and an end-to-end latency of 150 to 250 milliseconds.
In terms of pricing, Gemini Decision 1 charges $0.1 per million input tokens (about RMB 0.67), output tokens are free, and its cost is about one-fifth of the cost of calling Gemini 3.5 Flash-Lite through the Interactions API.
After the API launched, netizens had mixed evaluations. One user said: "After all that hype, this is it? Still, I will indeed use it for routing. Making a chat model specially generate a whole paragraph just to answer yes or no has always felt like overkill to me. But this is not exactly innovative either; we have seen quite a few similar things."
Another user raised a practical question: "If none of the given options are correct, how will it answer? Overall, I am glad to see this feature, especially since it supports multimodal input, which is useful for my asset classification work."
Another user mentioned that the value of this API depends on actual speed performance: "If it is really faster than having an ordinary model make the judgment directly, then that is interesting. I would rather get an answer in two seconds than wait for a long explanation I do not need at all."
On the same day, individual developer Lesia Kapliuk conducted a hands-on test using a free API key. Among the results, the model correctly classified all 54 calls, 30 customer service tickets, and 8 multilingual tasks, and processing 100 judgment questions in a single run took only 0.41 seconds. However, Lesia found that the model forcibly assigns inputs beyond the preset range to existing categories.
Currently, the Gemini Decisions API has been used in Google internal products such as Gemini Notebook, and Google Search also uses it for security protection of Model Context Protocol (MCP) tools. Developers can obtain an API key through Google AI Studio and use the free API quota for prototype development.
1. Supports mixing three types of judgment methods, processing up to 1,000 questions in a single request
The Gemini Decisions API processes developer-predefined questions through three types of judgment methods: Yes/No judgment, Multiple Choice, and Ordered Rubric, and these three methods can be mixed in the same request.
Yes/No judgment is used to estimate the probability that a certain condition holds, and developers can also separately set descriptive criteria for when the condition holds and when it does not hold. Multiple Choice supports 2 to 255 candidates; the model selects the option that best meets the requirements and returns the normalized probability corresponding to each candidate. Ordered Rubric evaluates the input content according to an ordered scale starting from 0, returning a weighted continuous score as well as the probability corresponding to each rating level.
All three judgment methods come with a confidence score between 0 and 1. Developers can set processing rules based on confidence, for example handing results the model cannot determine to a backup model, or escalating them to manual review.
In terms of multimodal processing, Gemini Decision 1 is based on Gemini's native multimodal architecture and can directly process text, images, audio, video, and PDF files in a single model computation. Different types of content can be combined in the same input, for example analyzing a call recording together with account information, or combining a PDF insurance claim document with on-site damage photos for judgment.
Developers can also adjust the resolution of images and videos to control the visual details the model needs to recognize and token consumption. A single request supports up to 256K tokens of input content, and up to 1,000 independent questions can be set for the same input.
Google also gave a practical application example. Developers can upload a photo of a breakroom countertop along with a text description, and have the model simultaneously judge whether there is uncleaned liquid on the countertop, which cleaning measure should be taken, and how dirty the scene is.
Among these, the model needs to judge whether there is a liquid stain, choose among three actions: wipe the countertop, arrange for cleaning personnel, and no action needed, and rate according to three levels: fully clean, minor localized stains, and severe stains or safety hazards. The three tasks can be completed in a single request.
Currently, multiple products within Google, including Gemini Notebook and Stitch, have already integrated the Gemini Decisions API, using it as a low-latency decision control layer. Developers only need to predefine the judgment options to let the model complete task assignment, without additionally training a dedicated classifier or waiting for the generative model to output a complete answer.
In addition, Google Search also uses the Decisions API as a security protection layer for Model Context Protocol (MCP) tools, helping to identify adversarial behavior and related content, improving detection effectiveness while controlling response latency.
2. Response latency as low as 50 milliseconds, cost less than one-fifth of generative models
In terms of operating mechanism, Gemini Decision 1 differs from traditional generative models.
Traditional generative models usually need to generate tokens one by one and then combine them into a complete answer. Gemini Decision 1 instead directly scores the candidate options predefined by developers, returning probabilities and confidence scores through a single forward model computation, without the need to generate answer tokens one by one.
Google says that in benchmarks covering agent routing, reasoning, ranking, and multimodal understanding, Gemini Decision 1 leads overall, and its calibration error is also the lowest among the APIs tested.
In multimodal input scenarios of up to 16K tokens, the model's median model computation latency is 50 to 100 milliseconds, and end-to-end latency is 150 to 250 milliseconds. When the number of questions in a single request increases from 1 to 100, the response latency remains basically stable.
In terms of cost, because shared input in the same request is billed only once, when adding judgment questions, users only need to pay additionally for the tokens consumed by the instructions and evaluation criteria corresponding to each question.
In benchmarks jointly conducted by both parties, compared with Gemini 3.5 Flash-Lite called through the Interactions API, the cost of Gemini Decision 1 with the Decisions API is nearly 80% lower, about one-fifth of Gemini 3.5 Flash-Lite.
Gemini 3.5 Flash-Lite charges $0.3 per million input tokens (about 2.01 yuan) and $2.5 per million output tokens (about 16.75 yuan).
Gemini Decision 1 charges $0.1 per million input tokens (about 0.67 yuan), and does not charge for output tokens, additional thinking tokens, or separate audio transcription and OCR text recognition fees.
In terms of call quotas, the Gemini Decisions API is divided into free and paid tiers.
The free version allows up to 200 requests per minute, processes up to 50,000 tokens per minute, and allows up to 10,000 requests per day.
The paid version charges $0.1 per million input tokens (about 0.67 yuan), with output tokens free. Under Tier 1, it allows up to 1,500 requests per minute, processes up to 1 million tokens per minute, and allows up to 50,000 requests per day.
3. Developer tests 54 calls, 100 questions increase time by 70%
Individual developer Lesia Kapliuk released the API test results of Gemini Decision 1 on November 11. The developer used a free API key to make 54 calls, testing the model's classification accuracy, multilingual processing capability, batch judgment efficiency, and result stability.
In actual tests, the developer submitted 30 customer service tickets and asked the model to assign the tickets to the corresponding teams; all 30 tickets were correctly classified. In multilingual tests, the model handled 8 tasks in languages including Ukrainian, German, Spanish, Polish, and Japanese, and also gave correct results for all of them.
In terms of batch processing, the developer tested cases with 1, 10, and 100 judgment questions in a single request. The results showed that when processing 1 question and 10 questions, the median response time was 0.24 seconds for both; after the number of questions increased to 100, the response time rose to 0.41 seconds, an increase of about 70%.
The corresponding token consumption also rose with the number of questions. Processing 1 question in a single request consumed 135 tokens, 10 questions consumed 613 tokens, and 100 questions consumed 5,473 tokens. In addition, when the same request was submitted 5 times repeatedly, the results and probabilities returned by the model were completely consistent.
This test also exposed a problem when the model handles abnormal input. When the input content does not meet the preset classification conditions, the model may still forcibly choose an existing option.
For example, the developer input the capital of Australia, meaningless random characters, and a passage requesting the creation of an autumn-themed poem, and asked the model to judge whether these contents were customer service requests.
After the developer added "Other" to the candidate options, the model chose "Other" for all three inputs, with corresponding confidence scores of 0.99, 0.69, and 0.98.
However, when the "Other" option was not set, the model incorrectly classified all three items into the technical support category, with confidence scores reaching 0.71, 0.64, and 0.84, respectively. For the yes/no judgment of whether they were customer service requests, the probabilities given by the model were only 0.01, 0.07, and 0.01.
It can be seen that although Gemini Decision 1 can quickly handle judgment tasks within the preset scope, its results are still affected by the design of the candidate options. If developers do not set a fallback option for abnormal input, the model may forcibly classify content that does not meet the classification conditions into existing categories.
Conclusion: Google extends model capabilities to fast decision-making scenarios
Judging from the positioning of the Gemini Decisions API, Google is trying to balance processing speed and call cost while remaining somewhat competitive.
For Agent applications that require frequent classification and tool selection, this type of API provides a new calling method. As Google internal products such as Gemini Notebook and Stitch are successively integrated, its application scenarios are also gradually expanding.
In the future, whether Gemini Decision 1 can continue to maintain stable judgment quality and processing efficiency in actual workflows will affect whether this type of API can form a long-term competitive advantage.
Source: Google