Google Launches Gemini Decision 1: Video and PDF Moderation for $0.10 per 1M Tokens

If you have user-generated content, you have a moderation queue. Videos, screenshots, voice messages, PDF contracts: all of it has to be checked against your rules. Running every file through a large general-purpose LLM is expensive and slow. Google is trying to close that gap with a dedicated product: Gemini Decision 1, a classifier model for fast decision assessment.

What this thing is

Decision 1 is not a chatbot and not an agent. It is a narrow tool: you give it material and a question such as "does this video violate the alcohol advertising rule?", and it returns a score. According to the description, the score is calibrated: it is not just a "yes/no" but a confidence value you can put a threshold on. Batch mode lets you ask up to 1,000 questions in a single request. The same file is checked against your entire rulebook at once, and you don't pay for re-uploading it.

Classifiers from Cloudflare and OpenAI already work on this logic. The author of the analysis in the Telegram channel @cgevent points to the key difference: Google's model can work with video. The model is natively multimodal, with a context window of 256,000 tokens. Inputs include text arrays, high-resolution images, audio, video streams and multi-page PDFs.

Why this matters in practice

Imagine a marketplace where sellers upload video reviews. The typical setup today looks like this: extract frames, run speech recognition separately, pass the text through one classifier and the images through another, then stitch the verdicts together. Each link breaks in its own way, and context is lost at the seams. If a neural network watches the whole video, part of that scaffolding can be thrown out. Google's models, starting with Flash, are already considered strong at understanding "pure" video without relying on subtitles or the audio track.

The second argument is money. The rate is $0.10 per million input tokens, with no surcharge for multimodality. According to Google, that is roughly ten times cheaper than general-purpose frontier models. For continuous auditing, where volumes are measured in millions of files, it is a matter of unit economics: checking everything rather than a sample.

The third argument is quality. In Google's own tests, the model beat TypeSafe's Jev 1 (the name as given in the source) on accuracy in difficult edge cases and on resistance to false positives. For moderation the latter matters as much as the former: every false positive is a blocked honest user and a support ticket.

What to check before adopting it

First, the caveats. The comparison with Jev 1 is Google's own data, and the analysis gives neither numbers nor a description of the methodology. The rate is quoted for input tokens, and the price of output tokens is not stated in the post, so check the documentation. Also keep in mind that video and audio are converted into tokens, so you should budget in tokens, not in "number of files". Long videos can eat a noticeable part of the savings.

What a developer or product manager should do:

• Build your own set of 200–500 labeled examples, especially edge cases: satire, quotes, borderline content. Only on such a set can you see the real false-positive rate.
• Run it through Decision 1 in Google AI Studio, where ready-made web demos are available, and compare it with your current solution on accuracy and cost per check.
• Use the calibrated scores rather than a binary verdict: automate the confident cases and send the uncertain ones to humans.
• If you pay for several narrow classifiers across different modalities, calculate whether a single call with a batch of questions could replace them.

The model is already available through the public Gemini API, so you can test the hypothesis in an evening. Most likely, the main win will not be in flashy benchmarks, but in video ceasing to be a separate, expensive branch of your moderation pipeline.

Source: Google's blog on the Gemini Decisions API (link in the primary sources of the @cgevent post).