Google's Gemini Decisions API Scores Up to 1,000 Questions per Request at $0.10 per Million Input Tokens. It Changes Yesterday's Self-Host Math.
Yesterday I wrote about Amazon engineers swapping Claude Sonnet 4.6 for a small self-hosted decision model and cutting ticket labeling cost by 98%. My break-even for running that model on a rented GPU was about 1,500 calls a day. A few hours later Google published the Gemini Decisions API, a hosted model that does the same kind of job and, per Google's blog, is "priced at $0.10 per 1M input tokens" with no output charge. That moves the break-even so far that most solo products should stop thinking about self-hosting a classifier at all.
What Google shipped
The model is gemini-decision-1, available on the Gemini Developer API through the v1beta endpoint. It doesn't generate text. Google's docs say "It generates zero text tokens, so output token usage is always 0." Instead, it answers closed questions of three types:
- Yes/no questions return a probability between 0.01 and 0.99, plus a confidence value.
- Choice questions pick from 2 to 255 options and return a probability for every option.
- Score questions grade against an ordered rubric of 2 to 255 levels and return a probability-weighted score.
One request can carry up to 1,000 questions against shared input, inside a 262,144-token context window. Input can be text, images, audio, video or PDF. The SDK shape is client.decisions.create(model="gemini-decision-1", input=..., questions={...}), and you read answers from result.answers[key].
The part I find most useful for small products: Google says the shared input is billed once per request, however many questions you ask. If you want to tag a support email with category, urgency, sentiment and "needs a human," that's four questions on one bill for the email's tokens.
The new break-even
Yesterday's numbers: an ml.g5.xlarge at $1.408 an hour is about $1,014 a month if you leave it running. The Amazon case study sent about 2,009 input tokens per ticket to its self-hosted model.
At $0.10 per million input tokens, a 2,000-token request costs $0.0002. Divide $1,014 by that and you get about 5 million requests a month, or roughly 169,000 a day, before the rented GPU is cheaper on paper. And that's with one question per request. If you ask four questions about the same input, the hosted side gets cheaper still while the GPU cost stays put.
Compare that to yesterday's 1,500 calls a day against Sonnet. Most solo products I know don't do 169,000 classifications a day. For them, the self-host option just lost its main reason to exist, unless the data can't leave your infrastructure.
What I couldn't confirm, and what it doesn't do
The $0.10 price is in Google's launch post. When I checked the official Gemini pricing pages, I couldn't find a Decision section yet. Neither the blog nor the docs say whether this is preview or generally available, and I found no region list. Google mentions a no-cost tier in AI Studio for prototyping. I'd confirm the billed price in your own console before building a budget on it.
The latency and benchmark claims are Google's own. The blog cites 150 to 250 ms end to end and says the model "leads across the suite, with the lowest calibration error," but the benchmark figures are in images, and nobody outside Google has reproduced them yet.
The docs also list what's off the table: no context caching, no Batch API, no function calling, no structured outputs, no grounding. If your current classifier relies on cached prompts to cut cost, that trick doesn't carry over. With zero output tokens and a low input price, it probably doesn't need to.
Where I could be wrong
The honest gap is the same one I flagged yesterday: accuracy. A cheap answer is only cheap if it's right. A model that returns calibrated probabilities is nicer to work with than one that returns a label and a guess, but I don't know how gemini-decision-1 does on your categories, and neither does Google. If your labels are domain-specific, a fine-tuned small model you control might still beat it, and then the cost comparison doesn't matter.
There's also vendor risk. This is a new API on a beta endpoint from a company that changes model lineups often. Keep the classification behind one function in your code so swapping providers is a one-file change.
What I'd actually do
Take 200 inputs you've already labeled by hand, from real traffic. Send them through the Decisions API as choice questions using the free AI Studio tier. Measure two things: how often the top choice matches your label, and how often a low-confidence answer flags the ones it got wrong. If agreement is high and the confidence values actually separate right from wrong, route your classification calls there and send only low-confidence items to a bigger model or to yourself. If it's worse than what you run now, you've lost an afternoon and a few cents of API calls.
Author
Lukas
@lukcombinator