Google launches Gemini Decision 1: decisions via API for $0.10 per million tokens

A customer writes: “My order arrived broken. I want a refund.” The application needs to identify the subject of the request and send it to the right employee. Google has launched Gemini Decision 1 for decision points like these: the model returns scores that code can use to choose the next action. A practical reason to take a look is the opportunity to automate request triage, content checks and other repetitive decisions within a product.

On Telegram, the new release was described as a competitor to Jev. That is an understandable starting point for discussion, but it is too early to choose a tool on that comparison alone: teams will still need to test quality on their own requests and calculate the cost of the workflow.

What developers get

The Decisions API documentation describes three modes: checking a condition, choosing from predefined options and rating on a scale. The response contains values suitable for processing in software, including probabilities and model confidence. No output text tokens are generated.

For a hypothetical store, developers could define separate checks: whether the customer is reporting damage, whether they are asking for a refund and whether there is enough information to process the request. Conventional code then combines the results. Suppose a request clearly concerns damage but does not include an order number. The next action could be to ask for the missing information.

It helps to distinguish between understanding a message and the system’s authority to act. Identifying the subject of a request is one task. Authorizing a refund is another, with its own rules. In this setup, the model helps interpret incoming requests, while the team defines the conditions for financial transactions separately. That makes it easier to identify exactly where an error occurred and fix the specific rule.

What a decision like this costs

Google’s official pricing lists $0.10 per million input tokens. Text, images, audio, video and PDFs are supported. Shared input within a request is billed once, regardless of the number of questions. However, each question’s instructions and criteria add their own tokens.

The phrase “at no additional charge” in the original post therefore needs to be read carefully: the absence of output tokens does not make additional instructions free. If a team sends lengthy criteria for every check, that volume must also be included in the calculation.

Take a hypothetical volume: ten million billable input tokens at this rate would cost one dollar. That is an arithmetic example of API costs. The full product budget will include integration development, result storage, failure handling and employees’ work on ambiguous requests. Product managers should calculate the cost per correctly processed request: this metric connects the model’s price to the business outcome.

A free tier is available. Google’s table states that data from this tier may be used to improve products; it states the opposite for the paid tier. This is a significant distinction when choosing data for a pilot.

How to test its value before deployment

Start with one clearly defined process, such as routing requests to departments. Gather historical requests with verified answers from employees and run them through the model. Identify ambiguous messages in advance: multiple problems in one email, sarcasm, incomplete information and requests that do not fit any department.

Measure incorrect assignments and the share of requests that need to be handed to a human separately. High model confidence alone does not establish quality on your dataset. For category selection, the documentation also warns that probabilities are distributed across the supplied options even when none is suitable. Include an “other” category and a route for manual review.

The next step is to run the system alongside employees while maintaining the existing workflow. This will reveal discrepancies before automated actions are enabled. You can then gradually entrust the system with the clearest cases and track how much time it actually saves.

Gemini Decision 1 merits this kind of pilot if your product already handles a stream of similar decisions. The case for deployment will emerge when testing demonstrates an acceptable error rate and lower processing costs.