Gemini Decision 1 vs Drex Max v1.1: Both Read 262K Tokens, Only One Counts Documents

Put G​emini Decision 1 and Drex Max v1.1 side by side and the specification sheets nearly overlap: both decision models, both a 262,144-token window, both reading images alongside text, both returning probabilities instead of prose. The gap opens in one place. On the document-decision component of the Artificial Analysis Decision Index v2.1 — the 30%-weighted test that asks a model to pull five decisions and ten facts out of a single dense document and be right about all fifteen — Drex Max v1.1 passes 63% of documents and G​emini Decision 1 passes 33%. That single component is most of the reason one scores 73 and the other 66.

The rest of the difference is arithmetic you can act on. Drex Max v1.1 costs $0.03 per million input tokens and G​emini Decision 1 costs $0.10; over a million tasks that is $103 against $345. And G​emini Decision 1 carries up to 1,000 questions per request over shared input, while Drex Max v1.1 expects one state with a handful of questions attached. Which of those numbers decides your choice depends entirely on whether your decisions come one document at a time or forty rules at a time.

Two different shapes of the same idea

Both models refuse to generate text, and both share one property that matters for cost modelling: the input is read once and every question is scored against it. Where they diverge is what they let you put in that input.

• Answer types — Drex Max v1.1 offers choice, noul and score, the request shape TypeSafe's J​ev defined and Nace adopted. G​emini Decision 1 offers the same three, with 2 to 255 options on a choice and 2 to 255 rubric levels, 2 to 10 recommended.

• Questions per call — up to 1,000 for G​emini Decision 1, each evaluated independently in one pass, keys you supply; a handful per state for Drex Max v1.1, where the unit of work is one item and its questions.

• Modality — G​emini Decision 1 accepts text, image, video, audio and PDF; Drex Max v1.1 accepts text and up to eight images per request. Neither accepts media inside the question instructions; on G​emini Decision 1 that is a documented limit, and there the media belongs in the shared input.

• Window — 262,144 tokens for G​emini Decision 1 across shared input and all questions together; Drex Max v1.1 documents 131,072 tokens per question with image tokens counted inside that, and a packed request ceiling of 262,144.

• Billing — G​emini Decision 1 charges $0.10 per million input tokens and explicitly re-bills repeated shared input in full, with no caching and no Batch API. Drex Max v1.1 charges $0.03 per million input tokens, a rate Nace halved when vv1.1 replaced v1 in place.

• Access — G​emini Decision 1 on the G​emini Developer API at v1beta, with a free tier and unlimited daily requests from Tier 2. Drex Max v1.1 hosted by Nace.AI at POST /v1/systemone, no downloadable checkpoint.

Read that list as one sentence: Drex Max v1.1 is built for a screen or a page and a few judgements about it, and G​emini Decision 1 is built for a big context and a large rulebook. The 1,000-question ceiling is not a spec-sheet flourish. It is the difference between a model you call forty times per request and one you call once.

Where the 6.6-point gap actually comes from

The Decision Index weights six evaluations, and both models have been run on all six. Lining up their components is more informative than the composite:

• AA-Paperclip, 30% — Drex Max v1.1 72.8, G​emini Decision 1 56.8. Fifteen answers per document, all-or-nothing. This is the sixteen-point gap that decides the ranking.

• EvalSafe-O*NET-AA, 15% — 88.3 against 87.1. A tie in practice.

• When2Call, 10% — 95.3 against 80.5. Drex Max v1.1 is second on the field here; G​emini Decision 1 is fourteenth of eighteen.

• AA-Omniscience-Decision v1.1, 15% — 60.3 against 66.3. G​emini Decision 1 takes this one, and its 44.8% raw accuracy is the highest of any decision model on the board.

• HLE-MultiChoice v1.1, 15% — 46.6 against 48.7. Close, both near the top, both with single-digit raw accuracy.

• AA-Symmetry v1.1, 15% — 79.8 against 70.5. Drex Max v1.1 keeps its answer under reordering and negation 53.3% of the time; G​emini Decision 1, 38.3%.

The pattern is consistent. Drex Max v1.1 wins the two components that test whether a decision is repeatable and structurally sound. G​emini Decision 1 wins the two that test whether it knows things and knows when it does not. Both are defensible products; they are answering different questions about what a decision model is for.

One caveat on the Drex figure: the $103 per million tasks that Artificial Analysis reports reflects a temporary launch price, and the board footnotes that the standard price would double the cost. At list, Drex Max v1.1 sits near $206 per million tasks — still well under G​emini Decision 1's $345, but no longer a third of it. Latency runs the other way: 283 milliseconds per task for Drex Max v1.1 against 211 for G​emini Decision 1, both slower than the field's 209 median.

Pick by the shape of your request, not the score

Take Drex Max v1.1 if your decision is about one artefact with an answer that has to hold. Insurance claim triage over a scanned form, screenshot-based agent control where the same click decision is re-asked after a re-render, contract review where a reversed yes/no is a real error. The 63% document pass rate and the 53.3% symmetry pass rate are the numbers that pay for themselves there, and $0.03 per million input tokens makes the volume argument easy.

Take G​emini Decision 1 if your decision is a rulebook applied to one input, or if the input is not text and images. Forty policy checks over a support ticket in one call is a workload G​emini Decision 1 was documented for and Drex Max v1.1 was not — and if the input is a call recording or a PDF, only one of these two can read it. G​oogle's docs list fourteen unsupported capabilities alongside the one it supports, so the fit is easy to verify before you commit: calibrated decisions, everything else off.

Take neither, yet, if the decision turns on cost at scale. G​emini Decision 1 has no Batch API and no context caching, which means a nightly sweep of a million historical tickets is billed at full rate with the policy document re-read every time. Drex Max v1.1's launch pricing is temporary. Neither product currently offers the discount structure that makes a backfill cheap, and that will matter more than six index points to anyone running the number on a year of tickets.

Running both without a second contract

OrcaRouter carries neither of these two. Both are probabilities-only endpoints rather than chat-completions targets, and we say so rather than implying otherwise — the catalogue's decision-model listing is TypeSafe's typesafe/jev-1.14, which sits at 66.23 on the same board between the two models in this piece.

What a bake-off actually costs you is the part that never appears in a benchmark: two SDKs, two key-management stories, two sets of rate limits to reason about. That is the reason the generative half of a decision pipeline belongs behind a router even when the scoring half cannot be. The call that drafts the candidate answer, writes the rubric, or produces the summary being scored is exactly the call where you want to swap vendors in an afternoon — one endpoint, more than 200 models, provider list price passed through at 0% markup, and failover when a provider degrades. Point the scoring call at G​oogle or Nace directly and keep the writing call portable. The model you cannot route is the cheap part; the model you can is the one you will want to change.

If you do test both on your own labels — and you should, because neither vendor publishes a calibration curve on your data — run them on the same 200 documents with the same question set and compare the pass rate, not the index. The board's 30-point gap on Paperclip is a strong prior. It is not a measurement of your documents.