Google's new API answers 1,000 yes/no questions about one PDF. A sub's submittal check is the obvious test, with one catch

Google released an API on November 11 that takes one document and up to 1,000 closed-ended questions and returns a scored answer to each. For a mechanical or electrical sub, the obvious use is a spec compliance checklist run against a submittal package before it goes to the GC. The catch: it tells you yes or no with a confidence number, and never tells you where it looked.

What did Google release?

The Gemini Decisions API runs on a new model, Gemini Decision 1, that does not write text. Per Google's post:

• Input: one shared input of text, images, audio, video or PDFs, up to 256,000 tokens.
• Questions: up to 1,000 independent questions per request. Each is a yes/no check, a multiple-choice pick from 2 to 255 options, or an ordered rubric score.
• Output: a confidence score from 0 to 1 on every answer.
• Price: $0.10 per million input tokens and nothing for output. The document is billed once, and each question adds only its own tokens.
• Access: the Gemini API and Google AI Studio, with a no-cost tier for prototyping. Google Cloud support is "coming soon."

At that price, a maximum-size input costs under 3 cents before the questions are counted. Google is not first here. The Decoder notes that a startup called Jev, then OpenAI, Cloudflare and Microsoft shipped similar models in recent weeks. We covered Microsoft's version and approval routing last month.

Why does this fit a submittal check?

A spec section is already a list of closed questions. Is the unit UL listed? Is the motor premium efficiency? Is the warranty five years on the compressor? Is the finish one of the three listed options? A submittal coordinator answers those by flipping between the spec and the cut sheets.

When we looked at Microsoft's model, the use was routing: one message, one pick. Google's pitch adds a different shape, with many questions against one document in one request. That is the shape of a submittal review, a closeout document check or a punch-list photo.

The last row is the one to be most careful with. The Decoder's example is a photo of a kitchen counter scored for a stain, a cleaning action and a severity level. That is a long way from judging a rated penetration.

What is the catch?

1. No reasons and no page numbers. Google says the model scores options "rather than generating answer text." A "no" on the warranty question does not point to the page. Someone still has to find it.
2. The accuracy claims are Google's. The post says Gemini Decision 1 has "the lowest calibration error" on its benchmark suite and gives no numbers. Nothing published covers construction documents.
3. The speed figures come from small inputs. Google quotes 150 to 250 milliseconds, tested with contexts up to 16,000 tokens. A full submittal package is far larger.
4. 256,000 tokens is a ceiling. A single product data submittal will often fit and a full project manual may not. Either way, the requirements have to be written as questions first. That is the real work.
5. It is an API. Nobody sells this as a submittal tool yet. A developer or your software vendor has to build it.

Should a sub act now?

Only if you have someone who can write a short script. If so, the test is cheap:

1. Take one closed-out submittal that came back "revise and resubmit," and the reviewer's comments.
2. Write the spec section's requirements as 30 to 50 yes/no questions.
3. Run the original package through and see whether the low-confidence and "no" answers land on what the reviewer caught.
4. Check the data terms of the tier you use before uploading anything an owner marked confidential.

If you do not have that person, ask your submittal software vendor whether they are testing decision models, and what they show the user when the answer is "no."

Takeaway: A cheap screen that sorts a submittal's requirements into "probably fine" and "look here" is worth testing, as a way to aim a coordinator's attention. It does not replace the coordinator or the engineer's review.

Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com and forward this to whoever assembles your submittals.