Google launches Gemini Decisions API to lower batch processing costs

Shared-input billing model

Google recently launched the Gemini Decisions API, a tool designed for closed-set decisions like policy checks, model routing, and context grading. A primary feature of this new endpoint is its unique billing structure, which charges for the shared input only once per request. This allows developers to bundle multiple independent questions into a single call without paying for the underlying data repeatedly.

As Tech Times reported, this approach can significantly lower costs for high-volume tasks. For example, performing 20 different checks on a 4,000-token record can cost about $0.50 per 1,000 records using this method. In contrast, sending those same 20 checks as separate requests—where the record is resent every time—would cost approximately $8.10 per 1,000 records.

This pricing model puts Google in a similar position to OpenAI, which also offers a shared-input structure for its Decisions API. Developers looking to optimize their workflows can now choose between vendors based on whether they prefer to pay per question or per request. Google also provides a no-cost tier in AI Studio for those who want to prototype their workflows before moving to a paid environment.

How the scoring works

Unlike standard language models that generate text one token at a time, Gemini Decision 1 functions by scoring candidate options in a single forward pass. Because the model does not generate any output text, Google does not charge for output tokens. This architectural choice is what enables the high speed and the specific pricing structure of the Decisions API.

The model is designed to handle up to 1,000 independent questions over a shared input that can reach 256K tokens. These inputs can be multimodal, meaning they can include a mix of text, images, audio, video, and PDFs. For every question asked, the model returns a score for the options provided and a confidence value between 0 and 1.

This confidence score is a critical feature for building reliable AI systems. It allows developers to write code that automatically flags low-confidence answers for human review or sends them to a more robust fallback model. For instance, if a model returns a confidence score of 0.92 for a specific answer, a developer can decide whether that is high enough to trust based on their specific safety requirements.

Accuracy and reliability risks

While the cost savings are substantial, the reliability of batch processing is a significant factor for developers to consider. Independent testing by Artificial Analysis showed that the leading decision model in their study passed only 63 percent of documents where all 15 required answers had to be correct. This suggests that as you add more questions to a single batch, the statistical likelihood of the entire batch being perfect decreases.

The math behind this is straightforward: every additional question represents another opportunity for the model to make a mistake. If a document requires 15 independent answers and each has a high individual accuracy, the probability of the entire set being correct can still be relatively low. This compounding effect means that high-volume batches require careful monitoring to ensure the data remains reliable.

Google claims that Gemini Decision 1 leads its suite in terms of lowest calibration error, which measures how well the model's confidence scores align with its actual accuracy. However, because Gemini Decision 1 was not included in the initial Artificial Analysis write-up, its specific performance on these complex documents is still being evaluated by the community. Developers are encouraged to run their own labeled records through the API to establish a baseline for their specific use cases.

Internal use and integrations

Google is already deploying this technology across its own internal products to improve performance and safety. For example, it serves as a real-time router in Gemini Notebook to send queries to the appropriate model tier, which Google claims can result in up to 40 percent lower latency. It is also used as a safety layer in Google Search to detect adversarial behavior in certain tools.

The API is also being tested in other areas, such as a project that generates playable wireframes in seconds and a tool that selects emoji reactions from a list of over 200 options. These diverse applications highlight the model's versatility in handling different types of decision-making tasks, from creative assistance to strict safety enforcement.

For external developers, the API is becoming increasingly accessible through various integrations. It is currently rolling out with popular platforms like LangChain, Vercel, and LiteLLM. Developers can access the API through Google AI Studio and can find documentation for Python, JavaScript, and REST to begin building their own decision-making workflows.