Google launches Gemini Decision 1: video and PDF decisions for $0.10 per 1M tokens

A familiar pain point: a product is full of small decisions that aren't worth a call to a large model. Where to send a request, whether to let a suspicious input through, which icon to pick for a response. Handing these to a flagship LLM is expensive and slow, and regex-based rules quickly become tedious. Google offers its own answer: Gemini Decision 1, a model from the class of decision models. Here is what is known about it from the announcement and from a review in the Tips AI Telegram channel.

What this model is and how it differs

The main difference, according to the review's author, is the range of input data. The model accepts text and images, as well as audio, video and PDF. According to him, neither OpenAI nor the Jev model offers such a set. The source does not explain what Jev is, so it is worth verifying this comparison yourself.

The second feature is batch processing. In a single request you can ask up to a thousand different questions about the same material, as long as everything fits into the 262K-token context. A simple example: you upload an hour-long call recording and a PDF contract, and in one call you ask a hundred things. Who named the deadline? Does the contract include penalties? Is the competitor mentioned more than once in the recording? Previously this was usually assembled from several services or run as separate requests.

Pricing and the free tier

Prices per 1M input tokens are as follows:

• Jev: $0.042;
• Gemini Decision: $0.10;
• OpenAI Decisions API: $0.10.

Google does not win on price: Jev is noticeably cheaper, and there is parity with OpenAI. The output price is not given in the post, so it is too early to calculate the full cost.

On the other hand, Gemini has a free tier. The limits are 200 requests per minute, 50 thousand tokens per minute and 10 thousand requests per day. The key is created in Google AI Studio. For Jev and the OpenAI Decisions API, according to the author, you have to pay right away.

One caveat from the numbers themselves. With 50 thousand tokens per minute on the free tier, you cannot load a 262K context in a single request: it will hit the per-minute limit. A thousand questions about a large document on the free tier most likely won't work. But many short requests, such as classification or routing, fit these limits well. That is the scenario for side projects.

Where Google already uses it

The author cites three internal cases, and they show well what the model is designed for:

• Gemini Notebook uses it as a router: for each request it chooses a suitable model and effort level.
• In Google Labs it picks an emoji for an AI agent from 200+ options.
• In Google Search it screens requests to MCP tools for attacks and harmful content.

All three tasks share one trait: the answer is short, the choice is made from a limited set, and latency and price matter more than eloquence. For developers of agent systems the third case is especially interesting. Almost everyone needs a filter between the agent and MCP tools, and it is usually built on homegrown rules or a heavy model.

What to do with it

If you have an agent or a chat with several models, try replacing an expensive router call with a model like this. Take a real request log, run it through and compare the choices with manual labeling. The same goes for a security layer for MCP: assemble a set of harmful and harmless examples and look at the share of misses and false positives.

What is missing from the source, and this matters: there are no quality benchmarks, and the verdict "not bad at all" remains the author's opinion. The output price, latencies and data-handling terms on the free tier are unknown. So there is no rush to put it into production with sensitive data, but connecting a key from AI Studio to a prototype in an evening is quite realistic.

The primary source is Google's blog on the Gemini Decisions API; check the current limits and prices there before adopting it.