Drex Max Is Out: the drex-max Model ID, Pricing and How to Call It

Short answer

NACE AI has published Drex, a family of four open decision models. Drex Max is the 27B member of that family, and its model ID is drex-max.

The release material is the paper "Towards Foundation Decision Models: Drex" (arXiv:2611.07854), submitted November 5, 2026 and revised to v2 on November 9. All four sizes landed at once: 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), and 27B (Max).

One thing to nail down first: Drex Max is not a chat model. A request has two parts — a state (a document, record or agent context, optionally with images) and a set of questions about it. Each question asks for a choice among options, a yes/no judgment, or a score on an ordered scale. The model returns a probability distribution over the options and does no autoregressive decoding. No token-by-token generation, and no output to parse.

The essentials:

• Parameters: 27B.
• Output: a probability distribution over options, a yes/no judgment, or a score on an ordered scale.
• Images: the state can carry images; in the paper, only Max received vision tuning.
• Open source: weights and code for all four models are published.
• API price: not announced.
• Context window: not announced.
• Typical latency: 183 ms, per the paper.

drex-max is not on TeamoRouter yet. Once it reaches our catalog, the key you already have works: change the model field to drex-max and keep the base URL. TeamoRouter carries no other NACE AI model, so there is no same-vendor stand-in to use in the meantime.

What problem it solves

Long chain-of-thought LLMs are the default because they reach higher accuracy across a wide range of tasks, but in production they come with three familiar problems: the same input gives inconsistent answers, the output needs parsing, and every step costs tokens and time.

A decision model removes that pipeline. You hand it the candidate options along with the question, it computes a probability for each one, and a single forward pass is the whole job. The engineering payoff is concrete: no parser to maintain, no format to drift, and latency that barely moves. The paper measures a typical Drex Max request at 183 ms.

The trade-off: it only answers multiple-choice questions. It does not write prose, generate code or improvise. It ranks the options you give it. Getting value out of Drex Max is about designing the option set, not about wording the prompt.

What the paper announced

Everything below comes from that paper:

• Four sizes: 8B (DLM), 9B (v1.5), 35B total with 3B active (v2.0), and 27B (Max).
• Request shape: a state plus a set of questions. The state can be a document, record or agent context and can carry images. Questions can be multiple-choice, yes/no, or a score on an ordered scale.
• No autoregressive decoding: it returns a probability distribution over the options directly, so nothing needs parsing.
• Training: supervised fine-tuning plus RLAF (reinforcement learning from adjusted feedback); Max also gets vision tuning.
• RLAF pays off: the ablations show it improves accuracy.
• Stable under reordering: consistency losses make the answer robust to the order in which options are presented.
• Fully open: weights and code for all four sizes are released.

The paper also reports Decision Index and Decision Bench numbers, plus head-to-head results against general-purpose models. This post covers the release, the specs and integration; the benchmarks belong in the comparison piece.

Specs at a glance

Pricing

There is no official API price to report this time. The release material has no per-token rate, no free allowance and no hosted API — the four models shipped as open weights, so the cost is whatever the hardware and time to run inference costs you.

So if you see "Drex Max API at $X per million tokens" somewhere, it did not come from this release. TeamoRouter pricing will be published when the model lands; there is no number yet.

Calling it (once it is live)

Not available yet. When drex-max enters our catalog, your key and base URL stay put — only model changes. OpenAI-compatible format:

from openai import OpenAI

client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
    model="drex-max",
    messages=[{
        "role": "user",
        "content": "state: <the document, record or agent context to judge>\n"
                   "question: does this record need human review? options: yes / no",
    }],
)
print(resp.choices[0].message.content)

Without an SDK, curl against the same host:

curl https://api.teamorouter.com/v1/chat/completions \
  -H "Authorization: Bearer $TEAMOROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"drex-max","messages":[{"role":"user","content":"state: <...>\nquestion: ... options: A / B"}]}'

Be clear about what the above shows: where model goes, not the full request contract. How state, questions and options are laid out inside a request — whether one call can carry several questions, how a score comes back — follows the interface docs after launch.

Vendor-native format: the release material defines no native HTTP endpoint or SDK for Drex. The models shipped as open weights plus code, so "native" today means loading them from the official repository yourself. There is no endpoint to copy here — rather than invent one, we will add the native path when NACE AI or TeamoRouter publishes it.

drex-max is not on TeamoRouter yet. Once it is live, the same key and model: drex-max is all you change; until then there is no same-vendor model on the site to use.

What you can use today

• No same-vendor substitute: none of the four NACE AI models is in our catalog yet, so there is no Drex edition you can start using on TeamoRouter now.
• Want to try the approach today: the other three members of the family are equally open — 8B (DLM), 9B (v1.5) and 35B with 3B active (v2.0) — and all can be self-hosted, with the smaller ones the lightest to run. Per the paper they follow the same approach as Max at different scales.
• If you need general chat or code generation: Drex is not that. It answers questions over options and does not produce free-form text. Reach for an existing general-purpose model for those.

How to pay

No subscription: top up your balance in the console from $5 and each request is billed by what it actually uses. Cards are supported, and so are USDT, USDC and ETH. The same key also calls Claude and GPT models.

New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.

FAQ

Is Drex Max out?

Yes. NACE AI published "Towards Foundation Decision Models: Drex" on November 5, 2026 and revised it to v2 on November 9, with weights and code for all four sizes.

What is the model ID?

drex-max. The paper calls this edition Max, and it is the 27B model.

How much does Drex Max cost?

No API price and no hosted API were announced. All four models are open weights, so the cost is your own inference bill. TeamoRouter pricing comes when the model is listed.

How long is the context window?

The release material gives no token count. It only says a request consists of a state plus a set of questions, where the state can be a document, record or agent context. How much text fits is a spec NACE AI has not published yet.

Can it write code or articles?

No. It is a decision model: it answers over the options, yes/no questions or ordered scales you supply, and it does not produce free-form text. Code and copy need a general-purpose model.

Does it handle images?

The paper mentions vision tuning for Max only, and the state can carry images. The other three sizes did not receive that training.

Can I call drex-max on TeamoRouter now?

Not yet — it has not entered our catalog. Once it does, the same key and the same base URL with a new model ID is all it takes.

Once it is live, is it just a model-name swap?

Yes. Key and base URL stay the same; only model becomes drex-max.

Related reading

• Getting access to Drex Max: accounts, balance and what to prepare before you actually call it
• Drex Max compared with other models: benchmarks, latency and where each one fits
• The paper: the official release material