Google Gemini

What is different about building on Gemini

Two things distinguish Gemini from a text-first API you may have used before.

Media is part of the request rather than a separate product. You attach an image, an audio clip, or a video to the same call you send text on, and the model reads them against your question.

And the unit of the API has changed. Google's current documentation is built around an Interactions API, where an Interaction represents "a complete turn in a conversation or task" and carries the execution steps, tool calls, and model outputs inside it. Google calls it "the best way to build with Gemini models and agents going forward."

That is a different shape from one request producing one completion. A turn can contain several steps, and the resource you work with holds them together.

Sending media in a request

Small media rides inline. Anything substantial goes through the Files API: you upload once, get a reference, and cite that reference in your calls.

The split matters more than it looks. Inlining a video means re-sending it with every message in a conversation, and paying for it every time. Uploading once and referencing gives you one transfer and one copy.

Media counts as tokens, and the rates differ by modality

Media is not an attachment. It is tokenized and counted against the same context window as your text, and the rates differ enough between modalities to change what you build.

As documented in September 2026:

Read the ratio rather than the numbers. Video costs roughly eight times what audio costs per second, so a ten-minute meeting recording is a manageable input as audio and an expensive one as video. If you only need what was said, send the audio.

Those rates will move. The durable part is the count_tokens method: call it with your request before you send it and you get the count for that model today, which beats any table including this one. The response also reports usage per category, so you can check what you spent.

Choosing between the Developer API and Google Cloud

Two front doors. The Gemini Developer API through AI Studio, and Google Cloud's platform for enterprise deployment.

The model capabilities are the same through both; the surrounding platform is not. Cloud brings the authentication, project structure, and governance controls of the rest of Google Cloud; the Developer API gets you an API key and a first call in minutes.

Google's documentation covers migrating between them rather than comparing them point by point, so treat the choice as an infrastructure question you already know how to answer. If your organization runs on Google Cloud and needs its controls, you will end up there. If you are building something small or trying the models out, start with the Developer API.

When Gemini is the right pick

Reach for it when media is central rather than incidental. A pipeline that reads scanned documents, answers questions about screen recordings, or works from audio has less glue code here than it would with a text model and separate services for each modality.

The token table above is the other reason. Handling long media is a budgeting problem, and a provider that tells you what a second of audio costs lets you plan for it.

Do not pick on benchmark scores you read this quarter. They move, competitors respond, and the cost of switching a well-structured application between providers is smaller than the cost of choosing wrong and staying. Keep your model call behind an interface and measure on your own task. See choosing the right model for how to run that comparison.

Further reading

• Google, Gemini API: the current surface, including the Interactions API.
• Google, Token counting: the rates above and the count_tokens method.
• Multimodal AI: the capability in general, across providers.