AI judgments just got cheap. Expect to buy far more of them
Published 2026-11-16
Jev and Google's Gemini Decision 1 bill input tokens only, at $0.042 and $0.10 per million. Cheaper judgments mean companies will make far more of them. How the money works, and a test to run on your own yes/no workload.
AI judgments just got cheap, and that will make companies spend more on them, not less.
Two models now bill input tokens only. TypeSafe's Jev charges $0.042 per million input tokens and does not meter output. Google's Gemini Decision 1 charges $0.10 per million input tokens and $0 for output. A yes/no answer, a label, a rubric score: each is now priced by what you send in, and nothing else.
This is the Jevons paradox. When something gets cheaper to use, people find so many new uses for it that total spending rises. The questions a team never sent to a model because the call cost too much (does this ticket qualify, does this invoice match, does this page pass review) are now worth asking. Expect far more of them, and a bigger total.
Teams are already moving. Within 24 hours on Vercel's AI Gateway, Jev reached nearly 13% of paid teams, twice the share of the GPT-5.6 family, according to The Next Web.
If your team sends yes/no or scoring questions to a full large language model (LLM), this issue is for you: what these calls cost, and a test to run on your own workload.
What happened
TypeSafe's Jev reached nearly 13% of Vercel AI Gateway's paid teams within 24 hours of arriving there. That was twice the share of the GPT-5.6 family. (The Next Web)
Google has a rival. Its Gemini Decisions API evaluates up to 1,000 yes/no, choice and rubric-scoring questions in a single pass. It works across text, images, audio, video and PDFs. (Google)
The money
Both vendors charge for what you send in, not for what comes back.
| Model | Input, per million tokens | Output |
|---|---|---|
| Jev (TypeSafe) | $0.042 | Not metered |
| Gemini Decision 1 (Google) | $0.10 | $0 |
Sources: The Next Web for Jev, Google for Gemini.
Google's input price is more than twice Jev's ($0.10 against $0.042). Neither one bills for output.
That matters because output can hide cost. Eden AI found that hidden reasoning output tokens made the same cheap model seven times more expensive on one route than on another.
Why it matters
If you buy AI for a team, your bill holds two kinds of work. One is writing, where a full LLM earns its price. The other is judging: does this ticket qualify, which of these four labels fit, does this document pass the rubric.
Judging is the half that just got cheap, and the half that will grow. Priced on input alone, its cost tracks how much you send and nothing else, so you can forecast it.
So the number to watch is not the price per call. It is the number of calls. When a judgment costs this little, your team will find more to judge.
What you can do with it
Who it is for: an operator or product lead whose team sends high-volume yes/no, choice or scoring questions to a full LLM, such as routing, screening and review.
The job: get the same answers at a cost you can forecast, and budget for asking more.
The move: pull that traffic out of your general LLM contract and price it on its own. Two per-input-token quotes are now public, $0.042 and $0.10 per million. Ask your current vendor to match one for judgment calls, rather than renegotiating the whole bill. Then budget for volume: the questions you skipped because they cost too much will fill the new line.
Founders have a separate bet to test. Google's Gemini Decisions API evaluates up to 1,000 questions in a single pass. Gemini Decision 1 bills $0 for output.
So a product built on it could price per request. Check what your route meters before you set that price, because output tokens can hide cost. No source gives Jev a per-request question limit.
The catch
Price is not quality. The sources give adoption and prices for Jev and Gemini Decision 1. They do not give accuracy against a full LLM on your questions, so a cheap answer that is wrong more often costs you more.
Adoption in 24 hours shows teams tried it. It does not show they kept it.
Cheap per call is not cheap in total. If the Jevons paradox holds, the saving on each judgment comes back as more judgments, so cap the volume, not only the price.
Try this
Take 1,000 real items from one yes/no or scoring workflow you now send to a full LLM. Google's Gemini Decisions API evaluates up to 1,000 questions in a single pass, so one batch fits there if that limit applies to the model you choose. The sources give no batch size for Jev, so confirm its per-request limit first.
- Decide what agreement rate with your current answers you will accept.
- Record what that workflow cost you last month, input and output tokens separately.
- Run the same 1,000 items through Jev or Gemini Decision 1.
- Compare the answers with your current ones, then compare the bill for the batch with your current cost for the same volume.
It worked if agreement clears your bar and the batch bill beats your current cost. Then list the questions you stopped asking because they cost too much: that list is where your spend goes next. If agreement misses, this workflow needs the full call.
Prices and adoption figures are as published by the linked sources. Confirm current terms with each vendor before you buy.
All source links
- thenextweb.com/news/typesafe-jev-decision-model-vercel-fastest-adopted checked 11 November 2026.
- blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api checked 11 November 2026.
- edenai.co/post/ultra-cheap-decision-models-sub-cent-ai-calls checked 11 November 2026.
