Gemini Decision 1 vs Jev: price, latency, media input and the thresholds a switch breaks

Gemini's confidence is a different number from Jev's confidence

Gemini's confidence is the probability of the top answer, while Jev's is a normalised spread that reads 0 when the options are evenly split. The same probabilities therefore produce two different confidence values, and a gate copied from a Jev integration passes more answers on Gemini. Jev also returns no confidence on yes or no answers; Gemini does.

The definitions, from the two docs pages on 13 November:

• Gemini: confidence "equals the top outcome's probability, or max(noul, 1 - noul) for yes/no questions" (docs).
• Jev: for a choice with n options, (p_max - 1/n) / (1 - 1/n); for a score, one minus the probability-weighted distance from the most likely level divided by that of an even spread, floored at 0; for a noul, no field (docs).

I wrote a script that applies both documented formulas to the same answers and checks the gates a routing config usually has. Part of the output, run on 13 November 2026 with Python 3.11.9:

answer                                    Jev  Gemini  gate >= 0.7 (Jev / Gemini)
choice, 4 options (0.61, 0.20, 0.12, 0.07)   0.48    0.61  fail / fail
choice, 4 options (0.75, 0.15, 0.06, 0.04)   0.67    0.75  fail / pass
choice, 3 options (0.80, 0.15, 0.05)     0.70    0.80  pass / pass
choice, 10 options (0.72, 0.28 / 9 each)   0.69    0.72  fail / pass
score, 3 levels (0, 0.57, 0.43)          0.35    0.57  fail / fail

gate               Jev 1.0    Jev 0.98     Jev 0.0  Gem 0.9234     Gem max  Gem 0.0498     Gem min
noul == 1.0          fires           -           -           -           -           -           -
noul > 0.99          fires           -           -           -           -           -           -
noul >= 0.95         fires       fires           -           -       fires           -           -
noul <= 0.0              -           -       fires           -           -           -           -
noul < 0.01              -           -       fires           -           -           -           -

"Gem max" and "Gem min" are the clamp ends, 0.99 and 0.01. The Gemini values in the middle are Willison's; the Jev values are the 1.0 and 0.0 he reported and 0.98 from TypeSafe's docs.

Two failures follow. A Jev gate written as noul == 1.0, > 0.99 or < 0.01 never fires on Gemini, so whatever it automated goes to the fallback path every time. A confidence >= 0.7 gate on a choice or score lets through answers Jev would have sent to review: the 10-option answer at 0.72 and the 4-option answer at 0.75 pass on Gemini and fail on Jev. Neither failure throws an error.

Of the four, only Gemini reads audio, video and PDFs

Gemini Decision 1 is the only one of the four major-vendor decision APIs that reads audio, video and PDFs directly, at the same $0.10 rate as text. Jev reads PNG and JPEG images only behind the TypeSafe-Danger-Zone-Images: true header and "does not natively accept audio or video (yet)". OpenAI Decisions takes images but no files or audio, and Microsoft-Decision-1 reads text and JSON.

Before 11 November, a Jev stack deciding about a voice call needed a transcription step and then a Jev call on the transcript. A decision about a scanned PDF meant OCR or page images through Jev's experimental flag. Google says the model "evaluates spoken words, pacing, and vocal tone without a separate transcription step" and reads "text, tables, charts, and page layouts together" in a PDF. That second claim is a capability a transcript cannot carry, which is why media is where Gemini earns the slot.

What media costs is less clear than the launch posts make it look. Google's docs give two audio rates, 25 and 32 tokens per second, on different pages, and none of them is specific to Gemini Decision 1. At those rates an hour of audio costs $0.009 to $0.012, a minute of video $0.0006 to $0.0018, and the 262,144-token window holds 2.3 to 2.9 hours of audio. Read the real count from usage.input_tokens_by_modality on your first calls. Inline uploads are capped at 100 MB per request, 50 MB for PDFs, and 15 MB per external URL; YouTube links and gs:// URIs are not accepted. Google's only media benchmarks are its own, 71.8% on MMAU audio and 69.1% on Video-MME, with no Jev or OpenAI score to compare. No one has published a timed comparison of a transcribe-then-Jev pipeline against Gemini reading the audio.

Switching from Jev to Gemini changes nine things, and two of them fail silently

Moving a Jev integration to the Gemini Decisions API changes the URL, the auth header, the model ID, one top-level field name, the yes or no criteria keys, how images are attached, the usage field names, and two numbers your code acts on. The question map, the choice criteria and the score criteria keep their shape.

The same three questions, as Gemini expects them:

curl -X POST "https://generativelanguage.googleapis.com/v1beta/decisions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-decision-1",
    "input": "I was charged twice for my subscription this month. Please refund the extra charge.",
    "questions": {
      "billing": { "type": "noul", "instructions": "Is this message about billing?",
                   "criteria": { "true_description": "About a charge, invoice or refund",
                                 "false_description": "About anything else" } },
      "tone": { "type": "choice", "instructions": "What is the tone of this message?",
                "criteria": { "angry": "An upset or hostile message",
                              "calm": "A neutral or polite message",
                              "excited": "An enthusiastic or eager message" } },
      "urgency": { "type": "score", "instructions": "How urgent is this message?",
                   "criteria": ["Can wait", "Needs attention this week", "Needs attention today"] }
    }
  }'

The last row is the pair that fails silently, as the gate test shows. The alphabetical sort also changes ties: exact ties break to the first key, which is why Google tells you to use "other" rather than "none_of_the_above". Questions of the box type have no Gemini equivalent.

A gateway hides most of the field changes. OpenRouter serves google/gemini-decision-1 on its decisions endpoint with the Jev-shaped body, state, questions and assets, though it marks that endpoint alpha. Vercel AI Gateway lists the same ID, and LiteLLM has gemini/gemini-decision-1 on its main branch and in a dev pre-release, but not in a stable release as of 13 November. No gateway changes the two numbers. The Microsoft-Decision-1 comparison and the OpenAI Decisions comparison cover those switches.

The case for starting a new build on Gemini Decision 1

The strongest argument against keeping Jev is that a team with no Jev integration pays no migration, and Gemini is the better-calibrated model on the largest independent index. It also packs 1,000 questions in one request and reads media, and its free tier covers a prototype. On short text, Duffin's run billed it the same as Jev.

That case holds for a team whose inputs are mostly media or long documents, whose calls run off the user's path, and who would set thresholds from scratch anyway. It fails on text at volume where latency counts: every independent test puts Jev at half Gemini's latency or less, and the Decision Index puts Gemini at 2.2 times Jev's cost per task. Gemini is also two days old, on a v1beta endpoint, without Vertex AI or a documented data region, which rules it out for teams whose data terms require either.

I would be wrong if an independent run showed Gemini at or below Jev's latency and cost per task on text workloads, or if TypeSafe moved audio and PDF input out of its experimental tier. On Stackness, one public profile lists Jev as of 13 November 2026 (mine), and none yet list the other three (data sources), so these are early calls in a small slot. The AI tools developers list in their stacks is where it will show when that changes.

Pick Jev if, pick Gemini Decision 1 if

Pick Jev if:

• your decisions read text or JSON
• each decision sits in a path where 100 ms matters
• your thresholds are tuned and you would rather not tune them again
• you need box questions or images inside question instructions

Pick Gemini Decision 1 if:

• the input is audio, video or a PDF and you would otherwise transcribe or OCR it first
• you ask hundreds of questions about the same input in one request
• your calls run in the background, where 200 to 500 ms is fine
• you start from scratch and will set thresholds on your own labelled data

The default for a stack that already runs Jev: keep it, and add Gemini as a second route for media calls only, with its own thresholds. Pick Microsoft-Decision-1 instead if you run on Azure and need inference in an Azure data zone.

Gemini Decision 1 in brief

• Free tier: yes, 200 requests a minute and 10,000 a day, with your content used to improve Google's products. Paid tier 1 allows 1,500 a minute (rate limits).
• Vertex AI: not yet; the launch post says Google Cloud support is coming soon.
• Runs locally: no. Strands Decider 2B is the local option in the slot.
• Batch and caching: neither is supported.