Nace Drex Max v1.1 ranking claim: what is confirmed so far

A claim says that Nace’s multimodal decision model, Drex Max v1.1, ranks first among decision models on Artificial Analysis’s leaderboard and sits on its cost-quality Pareto frontier. None of the sources the Unicorn News Desk reviewed show the model, that leaderboard or any statement from Nace, and none show who made the claim, so the ranking is unconfirmed.

This first article in a new thread sets out what the available evidence does and does not show, what the claim would mean if it holds, and what builders should check before relying on it.

What the Drex Max v1.1 claim says

The claim has three parts. First, Drex Max v1.1 is described as a multimodal decision model: a model that takes more than one kind of input, such as images and text, and is used to choose actions or outcomes rather than mainly to write prose. Second, it is said to rank first among decision models on Artificial Analysis, a third-party evaluation site. Third, it is said to sit on the cost-quality Pareto frontier, meaning no other model on that board is both cheaper and better.

No source reviewed contains a Nace announcement, a Drex Max model card, a technical report or a news report about the release. The sources also do not say who first stated the ranking. Until a primary record appears, every part of the claim, including the version number, the “decision model” category and the frontier placement, should be treated as unverified.

What the Artificial Analysis data in the sources shows

The only Artificial Analysis figures among the sources come through Benchmark Heaven, a site that republishes and filters leaderboard results. It shows the AA Intelligence Index, version 5.0.1, at a snapshot dated 2026-11-04, citing artificialanalysis.ai’s model leaderboard as the source. That is a general intelligence index, not a decision-model board.

The comparison on that page covers five models, all from OpenAI and Anthropic:

Model (configuration)	Lab	AA Intelligence Index
Claude Opus 5.5 (Max, Default Fallback)	Anthropic	59.1
GPT-6 Astra (Max)	OpenAI	56.9
GPT-6.1 Sol (Max)	OpenAI	54.6
Claude Sonnet 5.5 (Max, Default Fallback)	Anthropic	54.5
Claude Fable 5.1 (Max, Default Fallback)	Anthropic	49.2

Each score is marked as measured and observed on 2026-11-04. Benchmark Heaven also notes that the source “states neither a contamination control nor a task or question date window”. In plain terms, the page cannot say whether test questions might have leaked into models’ training data, or how recent the questions are.

Drex Max does not appear in this comparison. That is not evidence against the claim: the page compares a hand-picked set of five models, and the claim concerns a separate decision-model board that no source shows. It does mean the one Artificial Analysis snapshot available offers no support for it.

Where Drex Max is missing from other leaderboards

The desk also checked three other public ranking pages. Drex Max does not appear in the text of any of them.

• Onyx’s self-hosted leaderboard. This page ranks open-weight models that companies can run on their own hardware, comparing quality, speed, hardware needs and cost. It is credited to Roshan Desai and was last updated on 2026-10-13. Its top tier lists models such as Kimi K3, GLM-5.3, DeepSeek-V4.1-Flash and MiMo-V2.6-Pro. Drex Max is absent, which fits either a closed model or one not yet added; the sources do not say which.
• Concentrate.ai’s most-used rankings. This list orders models by usage. Its top five are OpenAI gpt-oss 120B, Claude Opus 5.5, Claude Haiku 5.5, OpenAI gpt-oss 20B and Claude Haiku 4.5. Drex Max is not in its top 50. A usage chart measures adoption, not quality, so it says nothing directly about the ranking claim.
• aiagent.org’s benchmark matrix. This page tracks models for coding, tool-calling agents and multi-step reasoning, and sets a minimum SWE-bench score of 40.0% for inclusion. The text available to the desk shows its methodology but no model rows, so it neither lists nor rules out Drex Max.

Why a cost-quality frontier position depends on the price maths

A Pareto frontier is a simple idea: plot every model by quality and by price, and keep only those that no rival beats on both at once. A model on the frontier is not necessarily the best or the cheapest; it is a sensible trade-off at its price.

The catch is that “price” is not one number. Model APIs charge separately for input tokens (what is sent in) and output tokens (what comes back). Benchmark Heaven lets readers switch between price bases, including output-only, input-only and fixed blends such as 1:1, 3:1, 10:1 and 100:1 input to output, plus an “adjusted costs” option and a setting that uses each model’s task workload as seen on OpenRouter. A model that sits on the frontier under one blend can fall off it under another, especially when its input and output prices differ sharply.

So even once a decision-model leaderboard showing Drex Max is published, a builder should check which price basis put it on the frontier and whether that matches their own traffic. A system that sends large images and short answers has a very different input/output mix from one that writes long plans.

How multimodal decision models are tested

“Decision model” is not a standard category on the general leaderboards in these sources, which group results into areas such as coding, agentic and tool use, and composite indices. How a decision-model board defines its tasks will shape any ranking on it, and no source describes the board the claim refers to.

Multimodal evaluation is also a crowded field. A public Hugging Face collection of multimodal benchmark papers, curated by the user btjhjeon, lists well over a hundred papers published between 2023 and 2025. With that many test suites available, a first-place result means little until it is clear which tasks were used, which input types were tested and how many competing models were included.

Scores also move with the testing setup. aiagent.org notes that results “may vary based on multi-sample voting, agent scaffold frameworks, and system prompt configurations”. Benchmark Heaven records the harness for many results, for example noting when a score is the best of Claude Code or Codex runs, or vendor-reported. A ranking claim should come with that context.

What it means for builders

For now, the practical answer is to wait for primary evidence. The ranking is an unattributed claim, and nothing in the sources shows how to access Drex Max: no API, app, open weights, licence, pricing or regions are documented. Builders evaluating decision or agent models today can only compare what the public boards actually list, such as the AA Intelligence Index snapshot above.

When more information appears, these are the checks that matter:

1. The source. A Nace post or model card, and the Artificial Analysis board itself, should both name Drex Max v1.1 and show the same result.
2. The field. How many models the decision-model board compares, and whether the frontier models from OpenAI and Anthropic that lead the general index are on it.
3. The price basis. Which input/output blend produced the frontier placement, given how much that choice can change results.
4. The test setup. Which tasks, input types and scaffolds were used, and whether contamination controls were stated.
5. Access and licence. Whether the model is available by API or as open weights, and on what terms.

A first place on a narrow category board can be useful if the category matches the job. It can also reflect a small field or a favourable price setting. Without the board itself, there is no way to tell which.

What to watch next

• A primary statement from Nace: a launch post, model card or technical report describing Drex Max v1.1, what changed from earlier versions and how it is accessed.
• A public Artificial Analysis decision-model leaderboard, or a page on artificialanalysis.ai, that lists Drex Max and shows its score, price and rank.
• Whether aggregators such as Benchmark Heaven pick up the result, with their notes on contamination and price basis.
• Whether Drex Max appears on usage charts such as Concentrate.ai’s or, if its weights are released, on Onyx’s self-hosted board.
• Independent tests from developers on real decision tasks, which would show whether a leaderboard position carries over to production work.

Unicorn API will update this thread if the claim is confirmed or withdrawn. For another recent look at an unverified model report, see Tencent’s DSpine fast-decoding AI checkpoint: what is confirmed so far.