Google Releases Gemini Decision 1: 150–250 ms Responses and $0.10 per 1M Tokens

It's a familiar pain: you're building an agent, and half of its calls to a large model aren't reasoning but routine. "Is this spam or not?", "which tool should I call?", "is there an error in this log?". Paying for text generation token by token just to get the answer "yes" or "option C" is wasteful and slow. Google has tackled this problem with a separate class of models.

What exactly was launched

Gemini Decision 1 is a "decision model." It works with a closed set of options: you provide a question and a list of answers, and it scores each option in a single pass. No text is generated token by token. Each answer comes with a calibrated confidence score.

Three modes are supported: yes/no, multiple choice (up to 255 options), and rubric-based scoring. Input can be text, images, audio, video, and PDF. A single request can hold up to 1,000 questions with a 256K context. End-to-end latency is 150–250 ms.

Price: $0.10 per million input tokens, with output not billed. Google calls this roughly 5 times cheaper than Flash-Lite. According to Phil Schmid, the project began as an internal DeepMind prototype, and a small team turned it into a model and API in about six weeks.

What it looks like in practice

Imagine a support team receiving thousands of tickets an hour. Previously, you asked an LLM to "determine the category and urgency and write a response in JSON." Now you send a batch of questions like "is this a billing complaint?", "is the customer angry?", "is a human needed?" and get answers with confidence scores. If confidence falls below a threshold, the ticket goes to a human or to a large model.

Cloudflare demonstrated this setup at AI Engineer. Fast decision models choose tools and pull the relevant lines from logs, while a large reasoning model is brought in only for diagnosis. In a demo of a failed build, six LLM calls became one, and consumption dropped from roughly 40K tokens to 1.8K. LangChain added support for Decision 1 on release day, with a shared base class for TypeSafe classifiers, so connecting it to an existing pipeline is not hard.

What independent tests say

Google's own benchmarks look convincing: Decision 1 beats competitor Jev 1.14 from TypeSafe in 5 of 7 shared tests, for example 92.2% versus 90.0% on JevBench. But Artificial Analysis cools the picture.

• In the aggregate Decision Index, the model scored 66 points. That is on par with Elephant 9B and Jev 1.14 and 7 points below Drex Max v1.1 and GLiDE v1.2.
• It leads in AA-Omniscience-Decision (66) and HLE-MultiChoice (49).
• In AA-Paperclip, the model answers every question flawlessly for only 33% of documents; Drex Max scores 63%.
• The price per 1M tasks is $345 versus $103 for Drex Max and $156 for Jev.
• Average time per task is 211 ms; for Jev it is 92 ms.

A cheap input rate doesn't mean a cheap task: by AA's calculation, the cost of real work is higher than for several competitors. And on speed, Jev is almost twice as fast. Gemini Decision 1's advantage is multimodality and a large window of 1,000 questions per request, not absolute quality.

What to do with this

Developers. Find the calls in your agent where the answer is a choice from a finite list: routing, tool selection, filters, log extraction. These are worth running through a decision model, leaving the expensive LLM for real reasoning. Confidence thresholds will help decide when to escalate.

Product managers. Calculate the cost per task, not the price per token. Run your own dataset through Decision 1, Jev, and Drex and compare accuracy, latency, and total cost on your data. Vendor benchmarks are no help here, as always.

Founders. The market for "decision" models is only taking shape, and Google entered it at the same time as TypeSafe, which already has paying customers. Expect a price war and the quick arrival of new competitors.

The limitations are clear: the model won't replace generation; it is good on closed sets of options but noticeably trails the best competitors on complex documents. The data so far comes mostly from Google and one independent lab, so it's best to make a decision after running your own test.