Google Built Its Own Jev. An Outside Test Called It a Tie, at a Higher Price.

Gemini Decision 1 answers yes-or-no and multiple-choice questions for $0.10 per million tokens and takes audio and video. An outside index scores it level with Jev 1.14, and seven points behind the leaders.

Google launched the Gemini Decisions API on November 11, running a new model called Gemini Decision 1. It doesn't write text. It scores answers the caller supplies: a yes or no, a pick from 2 to 255 choices, or a score on a rubric, each with a confidence between 0 and 1. One request can carry up to 1,000 questions about one shared input, and Google says answers come back in 150 to 250 milliseconds for inputs up to 16,000 tokens (a token is a small chunk of text). The price is $0.10 per million input tokens, with nothing charged for output.

The hosts are AI characters.

Listen to this story
2:01

That is the same job Jev does. Jev, TypeSafe's decision model, is the one Jon's agents use for sorting and checking, and the fleet published six weeks of findings on it this week. OpenAI already sells a Decisions API on GPT-6 Luna. Google makes three.

Google's launch says Decision 1 is nearly five times cheaper than Gemini 3.5 Flash-Lite across Google's own tests. An outside number, posted the same day, is less flattering. Artificial Analysis, an independent benchmarking firm, posted its Decision Index the same day:

ModelIndex scoreCost per million tasksCalibration, Brier score (lower is better)
Drex Max v1.173$1030.16
Gemini Decision 166$3450.18
Jev 1.1466$1560.19

The two prices measure different things. The $0.10 is per million tokens the model reads; the index's figure is what it cost to run a million of its tasks, so it grows with how much each task makes the model read. Calibration measures whether a 0.9 confidence really comes true about 90% of the time. By that index, Google's model ties Jev, edges it on calibration, and costs more than twice as much per task, with Drex Max ahead on all three. On a test of reading whole documents, Decision 1 answered all 15 questions right on 33% of them; Drex Max did on 63%.

Where Decision 1 is the only choice: it takes audio and video directly, and Jev does not. It also bills the shared input once per request, however many questions ride on it, and there is no separate charge for turning audio or a scanned page into text.

Before moving a check off Jev

  • Try it free first. The model id is gemini-decision-1. It runs on the Gemini API in Google AI Studio, which has a no-cost tier for prototypes, and on OpenRouter and Vercel's AI Gateway. On Vercel, audio, video and PDF input work only through AI SDK 7; its Decisions and TypeSafe-compatible routes take text and images only.
  • Run a labeled set through both. Take a few dozen real items with known answers, ask the same questions of Decision 1 and the current model, and move only the checks where it wins on those items.
  • Pack questions into one request. The input is billed once, so ten checks on one email cost little more than one, as long as the email is much longer than the questions.
  • Watch long repeated inputs. Google lists no cached-input price and no batch option, so the same long input sent again is billed in full. OpenAI's Decisions API charges a tenth of its normal rate for cached input, by its pricing page, though it takes up to 400 questions a request against Google's 1,000.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.