Introducing Drex for Finance

Most of a fintech stack runs on small decisions. Is this invoice a duplicate? Is this login an account takeover? Which risk band does this applicant fall in? Is this Fed statement hawkish or dovish? Each one has a few possible answers, and each one needs an answer fast.

Today we're releasing Drex for Finance, a version of Drex Max trained further on financial decisions. You send it a case, such as a transaction, an invoice, a support message or a filing excerpt, with the questions you want answered and the options for each. It picks an option and tells you how sure it is. It can only answer with the options you list, so it can't make up an answer.

On 48 public finance test sets it never saw in training, Drex for Finance scores 77.3 on the Drex Finance Index. Jev 1.14 scores 63.6, GPT-6 Luna 58.4 and pplx-decider v1.2 65.0. It gets 85.7% of 3,361 decisions right, and a short case comes back in a median 0.82 s.

Benchmark	Drex for FinanceNace	Jev 1.14TypeSafe	GPT-6 LunaOpenAI	pplx-decider v1.2Perplexity	Drex Max v1.1Nace	Drex Max v1Nace	GLiDE
Overall
Drex Finance Index	77.3	63.6	58.4	65.0	65.8	60.1	65.4
Accuracy, %	85.7	76.7	70.9	77.4	77.7	73.1	77.6
Language & signals
Sentiment & tone	80.7	69.1	68.6	73.9	69.3	58.6	73.0
News & headlines	89.5	66.4	60.7	68.8	71.4	68.3	68.4
Banking & support intent	87.1	71.6	71.9	73.7	74.5	65.6	73.9
Documents & filings
Documents & extraction	92.3	85.9	48.8	87.1	87.8	67.2	88.8
Financial QA & filings	96.7	86.4	88.1	82.6	89.7	84.5	89.1
Risk, macro & markets
Credit & risk	48.3	30.0	38.3	11.7	0.0	21.7	0.0
Monetary policy	53.8	45.0	38.8	35.0	30.0	25.0	33.8
ESG & climate	73.0	59.1	48.1	54.4	59.4	60.0	56.1
Markets & reasoning	41.5	28.1	30.1	43.1	43.0	40.1	41.4

Each of the 48 test sets is scored on skill above chance, and the index combines all of them. A model that does worse than a coin flip on a set gets 0 there, not a negative score.

Results by category

Drex for Finance has the best score in 10 of the 11 rows. Against Jev 1.14 it's ahead in all nine categories, by the most in News & headlines (+23.1), Credit & risk (+18.3) and Banking & support intent (+15.5). The closest is Documents & extraction, at 92.3 against 85.9.

Two rows are worth a closer look:

• Credit & risk. Drex Max v1.1 and GLiDE score 0.0. That's the floor, not missing data: both flag default and bad credit far too often, which puts them at or below a coin flip. Drex for Finance scores 48.3, 10 points ahead of the next model.
• Markets & reasoning is the one category where we're not first. Drex for Finance scores 41.5, behind pplx-decider v1.2 at 43.1 and Drex Max v1.1 at 43.0. The finance training didn't move it.

Results on general decision benchmarks

A model trained hard on one domain often slips everywhere else. This one didn't. On our general decision benchmarks, JevBench went from 93.9% to 94.4% and Decision Bench from 94.2% to 94.6%.

What it is trained on

The training data covers nine areas:

• Banking & payments: customer intent and routing, card, transfer and dispute triage, chargeback reason codes, payment failure causes, fraud and phishing flags, account-takeover signals.
• Accounting & audit: invoice and receipt fields, expense policy checks, GL account classification, reconciliation exceptions, duplicate invoices, audit sampling and risk flags, revenue recognition checks.
• Credit & lending: applicant risk bands, default flags, affordability checks, loan document completeness, covenant breaches, collections priority.
• Insurance: claim routing, coverage eligibility, policy questions, claim fraud indicators, underwriting risk tiers, subrogation flags.
• Compliance & risk: KYC and AML red flags, sanctions and PEP screening, adverse media, disclosure checks, transaction monitoring alerts, regulatory change relevance.
• Filings & reports: 10-K, 10-Q and XBRL tagging, numeric claim checks, table and text QA, risk factor classes, MD&A tone, footnote extraction.
• Markets & investing: news, social and entity sentiment, monetary-policy stance, M&A and event signals, earnings call tone, analyst rating direction, trade signal gating.
• ESG & sustainability: ESG claims, climate risk disclosures, greenwashing flags, taxonomy alignment, controversy screening.
• Treasury & operations: cash forecasting signals, vendor risk, procurement approvals.

What you can build

Drex for Finance fits anywhere a pipeline needs a fast, scored answer from a fixed set of options:

• Documents & filings: tag and classify filings, check numeric claims, code invoices to GL accounts, flag reconciliation exceptions and duplicate invoices.
• Credit & lending: assign applicant risk bands, flag likely defaults and covenant breaches, rank collections, check that a loan file is complete.
• Fraud & AML: score transactions before they clear, catch account-takeover signals, triage disputes and chargebacks, screen names against sanctions and PEP lists.
• Markets & trading: read sentiment from news and earnings calls, classify central bank statements, label events, and gate trade signals before they fire.
• Insurance & regtech: route claims, check coverage, flag claim fraud and greenwashing, and decide which regulatory changes matter.

Because every answer is a probability, you choose where the line is. Approve or flag the clear cases automatically, send the uncertain ones to an analyst, and log the score next to each call.

Get started

The model id is drex-max-finance, and a request can be up to 262,144 tokens. There are three ways in, from no setup to production.

Try it in your coding agent

Add Nace's remote MCP server to Claude Code, Codex, Cursor or any MCP client, then ask your agent to make a decision with drex-max-finance:

claude mcp add --transport http drex https://console.nace.ai/mcp

For Codex, run codex mcp add drex --url https://console.nace.ai/mcp. Without a key, the free tier allows 100 decisions a day of up to 32 KB each. Add a key when you need more.

Call it with the SDK or CLI

Sign in at the Console, create an API key and export it as NACE_API_KEY. Eligible new accounts get $5 in free credits. Then install the SDK for Python or TypeScript:

# pip install nace-sdk
from nace_sdk import Choice, NaceClient, Noul, Score

invoice = (
    "INV-2304, Northwind, $9,870.00, PO-77120. "
    "Paid Oct 12: INV-2304A, Northwind, $9,870.00."
)
gl_accounts = {
    "5100": "Raw materials",
    "6200": "Office supplies",
    "6400": "Professional services",
}
questions = {
    "duplicate": Noul(
        instructions="Is it a duplicate of a paid invoice?"
    ),
    "gl_account": Choice(
        instructions="Which GL account is it?",
        criteria=gl_accounts,
    ),
    "risk": Score(
        instructions="How risky is it to pay?",
        criteria=["low", "medium", "high"],
    ),
}

# reads NACE_API_KEY from the environment
with NaceClient(model="drex-max-finance") as client:
    result = client.system_one(
        state=invoice, questions=questions
    )

print(result.nouls["duplicate"].noul)       # P(yes)
print(result.choices["gl_account"].choice)  # a label
print(result.scores["risk"].score)          # 0 to 2

All three questions share one forward pass and one bill. npm install nace-sdk gives the same client in TypeScript, and the nace CLI runs the same call from a shell:

pipx install nace-cli && nace login
nace decide -m drex-max-finance --state @invoice.txt \
  --noul "duplicate=Is it a duplicate of a paid invoice?"

Switch from Jev

The Drex API is wire-compatible with TypeSafe's Jev. Keep @typesafe-ai/sdk and change three environment variables:

TYPESAFE_BASE_URL=https://console.nace.ai
TYPESAFE_API_KEY=nace_sk_...
TYPESAFE_DEFAULT_MODEL=drex-max-finance

Code written for OpenAI's Decisions API works too: point the OpenAI SDK at https://console.nace.ai/v1 and set the model to drex-max-finance.

Drex for Finance is available today in the Nace Console. The use cases and category scores are on the Drex for Finance page, and the docs cover question types, limits and errors.