Drex for Finance vs Gemini Decision 1: Two Ways to Buy a Probability

Put the spec sheets side by side and Drex for Finance and G​emini Decision 1 look like the same product from two companies. Both take a block of input and a set of typed questions, both return a scored answer per question with no generated text, both cover a 262,144-token request, and both cost cents per million input tokens rather than dollars. The resemblance stops at the thing you would actually buy them for. G​emini Decision 1, released 11 November 2026 on G​oogle's Developer API, is a general decision model whose distinguishing skill is knowing when to decline; Drex for Finance, generally available since 2 November 2026 at $0.03 per million input tokens, is a domain finetune whose distinguishing skill is reading a financial document and getting fifteen answers right at once. Nine days apart, same request shape, opposite bets — and no published run has ever put them on the same scale.

The comparison nobody has run

This is the first thing to understand, because it determines how much of the rest of this piece you can act on.

G​emini Decision 1's numbers come from Artificial Analysis, on its Decision Index v2.1 — eighteen models ranked, a weighted composite of six evaluations, each scored as accuracy minus the Brier penalty on the probability the model attached to its answer. It scores 66.0, which is ninth of eighteen, just above the board median of 64.

Drex for Finance's numbers come from Nace, on a Drex Finance Index built from 48 held-out finance test sets. It scores 77.3. Drex for Finance is not on the Artificial Analysis board at all.

So the two headline numbers are 66 and 77.3, they are on different instruments, they measure different populations of tasks, and subtracting them produces a number with no meaning. Anyone who does it — and the versions of this comparison already online do — is comparing a general-purpose board score to a vendor's own domain score and calling the difference a winner. The honest statement is that the only overlapping evidence is one step removed: Drex Max v1.1, the parent checkpoint Drex for Finance was finetuned from, does sit on the Artificial Analysis board, and it leads it outright at 73.4 — first of eighteen, 7.4 points clear of G​emini Decision 1's 66.0. That is a strong signal about the base model and no signal at all about the finetune, because domain training can gain a lot on a document test and lose a little on a knowledge one, and nothing published separates those two effects for this checkpoint. It is also the reason the 66 and the 77.3 must not be read as a gap in either direction.

What each one is actually built to do

The architectures explain the divergence better than any score does.

• Shape — Drex for Finance is a finetune of Drex Max, a decision model with a vision encoder. G​emini Decision 1 is a first-of-line decision model from G​oogle DeepMind, on its own POST /v1beta/decisions endpoint, separate from the generate-content path.

• Training emphasis — nine named finance domains at Nace: banking and payments, accounting and audit, credit and lending, insurance, compliance and risk, filings and reports, markets and investing, ESG, treasury and operations. G​oogle publishes no training description beyond the model and the API.

• Scale of a request — both allow 1,000-ish questions in one pass, but the shapes differ. G​emini Decision 1 caps at 1,000 questions per request and bills the shared input once. Drex for Finance gives you 262,144 tokens across the whole request with the question count unbounded in the documentation, which rewards many short questions over one long document.

• Abstention — G​emini Decision 1's yes/no type returns a probability bounded to 0.01–0.99, so a threshold set above 0.99 can never fire; its board leadership on AA-Omniscience and HLE-MultiChoice is substantially a leadership in declining rather than guessing. Drex for Finance returns P(true) in [0,1] with no documented bound, and no published non-hallucination rate at all.

• Option handling — G​emini Decision 1 sorts your option keys alphabetically in UTF-8 byte order before scoring, so ties break toward the alphabetically first key, and G​oogle's own guidance is to name a catch-all other rather than none_of_the_above. Nace documents no such normalisation, which is not the same as saying it does not happen — it is the same as saying you cannot rely on it either way.

• Price — Drex for Finance $0.03 per million input tokens, output not billed. G​emini Decision 1 $0.10 per million input tokens, output not applicable. On Artificial Analysis's task-rate measure G​emini Decision 1 comes to $345 per million decisions; the comparable figure for the Nace model is not published on that basis.

• Access — Drex for Finance through the vendor's API, its SDKs, its CLI and a remote MCP server, with a free tier of 100 decisions a day up to 32 KB each. G​emini Decision 1 on the G​emini Developer API at v1beta only — no Vertex AI listing, no Batch API, no context caching.

Where the general model wins, and it is not a small set

Three of those rows are decisive depending on your workload.

Calibration. The Decision Index penalises a confident wrong answer harder than an uncertain one, and that design is why G​emini Decision 1 sits where it does. If your pipeline routes on the probability — approve above 0.95, escalate below 0.6 — the value of the number is entirely in whether 0.95 means 95%. G​oogle's bounded output plus its first-place finish on the two knowledge components is evidence for that; Nace publishes accuracy and a category breakdown but no calibration curve, so the Drex probability is a score you have to validate yourself before you gate on it.

One pass, many rules. A thousand questions per request against one shared input is a genuinely different capability from a very long single context. A support ticket checked against forty policy rules in one call, billed once for the input, is the kind of job where G​emini Decision 1's shape is better than a 262K context window, because the cost scales with questions rather than with the text you repeat.

Modality breadth. Drex for Finance reads text plus PNG, JPEG and WebP images, up to eight per request. G​emini Decision 1 takes text, image, video, audio and PDF in one request, up to 100 MB of inline base64, 50 MB for PDFs. If a decision depends on a phone call recording or a video frame, only one of these two can see it.

Where the specialist wins

The counter-case is narrower but it is real, and it comes down to what happens on documents that need several answers at once to all be right.

On AA-Paperclip — 30% of G​emini Decision 1's composite, and the heaviest single component — G​emini Decision 1 passes 33% of documents outright, against 63% for Drex Max v1.1. The test is 1,500 private resumes, receipts, invoices and papers, each carrying five decisions and ten facts, scored all-or-nothing per document. Drex for Finance was finetuned on exactly that material: invoice and receipt fields, GL account classification, duplicate invoices, reconciliation exceptions, audit sampling, 10-K and XBRL tagging. Nace reports 92.3 on Documents & extraction and 96.7 on Financial QA & filings, both vendor-run.

If your decision is "read this invoice and tell me whether it duplicates one already paid", the general board's composite is close to irrelevant and the domain finetune is close to the whole answer. That is the case for paying a specialist rate for a specialist model — and it is also why the 77.3 and the 66 should not be compared: they are the scores of two models on the two task populations each was chosen to be good at.

Two honest weak spots on the Nace side, since the table is the vendor's: Credit & risk scores 48.3, the lowest absolute number in the finance table, and Markets & reasoning scores 41.5, the only category where the model is not first — pplx-decider v1.2 takes 43.1. Both are disclosed in Nace's own launch post, which is more than most vendors do, and both bound what the finance checkpoint should be trusted with.

Cost, at the scale a decision pipeline actually runs

A 4,000-token case scored once costs about $0.00012 on Drex for Finance and about $0.0004 on G​emini Decision 1. Neither is the number that decides anything; at a million decisions a month the difference is roughly $280, which is less than the engineer-week spent migrating.

What does decide it is the surrounding architecture. G​emini Decision 1 has no context caching and no Batch API, so the same 40,000-token policy document is billed in full on every request, and there is no half-price overnight sweep of historical tickets. Drex for Finance has neither feature either, and its billing is per request rather than per question, which means the cost of asking forty things about one document is the token count of the document, once. On a high-question, short-input workload that inverts the per-token ranking entirely: G​emini Decision 1's $0.10 buys you a thousand questions, while the same thousand questions against one short case is one Drex request at $0.03 per million tokens.

How to reach either one, honestly

We do not carry gemini-decision-1 and we do not carry drex-max-finance. A probabilities-only endpoint is not a chat-completions target, so a router in our shape has no honest way to put either behind one key, and claiming otherwise would fail at the first request. What we do carry is TypeSafe's typesafe/jev-1.14 at $0.042 per million input and output tokens on systemone — the same request shape Nace's API is wire-compatible with and a model on the same eighteen-model board as G​emini Decision 1, at 66.2 against 66.0.

Which makes it the practical third option in this comparison rather than a footnote. If the decision itself is uncontested and what you need is one endpoint for the whole pipeline — the decision call, the drafting call that writes the analyst note, the escalation call that reviews the flagged cases — then the decision leg running on the same key as more than 200 other models, at provider list price passed through with no per-token markup, is worth more than a two-point edge on a composite. If the decision is the whole product, buy the specialist from the vendor and route the rest.

What would settle it

One experiment, and nobody has run it. Take a few hundred real finance documents — invoices with known duplicate status, filings with known tags — and score them with both models on the same rubric. That single run would replace every number in this article with a comparable one and would tell you whether the domain finetune's advantage survives contact with an independent harness.

Until it exists, the correct reading is that these are not competitors. G​emini Decision 1 is the better choice when the probability has to be trustworthy, the questions are many, or the input is not text. Drex for Finance is the better choice when the input is a finance document, the answers have to all be right at once, and the case volume makes $0.03 per million tokens the line that matters. Both statements are defensible from published numbers. The comparison that puts them on one leaderboard is not, and it will be a while before one exists.