Jev vs Drex Max: a higher-accuracy System One model at $0.06 per million
Nace.AI’s Drex Max is on OpenRouter as nace-ai/drex-max. It takes the same System One request as Jev — a state plus typed Choice, Score and Noul questions — and returns calibrated probabilities without generating text. In our run on 2026-11-20 it beat Jev on accuracy (91.7% against 86.5% on Banking77 card intents, 94% against 90% on AG News), used fewer tokens despite a higher list price ($0.06 against $0.042 per million), and was slower. On JevStation both cost 1 credit standard and 3 large.
Try it on your own text
Edit the example, then sign up free to run it — 15 credits, no card.
Scenario
Route a ticket to the right team, rate its urgency and flag churn risk.
Edit freely — the model answers the scenario's questions about whatever you put here.
What Jev actually returned
This is the unedited output of the example above, captured from the live API.
Hi, I've been trying to connect my Stripe account for 3 days and it keeps failing with 'invalid redirect'. I'm losing sales every hour this is down. Please help ASAP.
category
choiceWhich team should handle this support ticket?
integrations
- integrations99%
- billing1%
- bug0%
- account0%
Confidence 99%
urgency
scoreHow urgent is this ticket?
- 4 · Critical73%
- 3 · Urgent27%
- 0 · Can wait0%
- 1 · Normal0%
- 2 · Soon0%
Confidence 77%
churn risk
noulIs the customer at risk of churning?
Captured 2026-10-09 · 0.89 s · jev-1.13.0
At a glance
(vendor) is the company selling the model; (ours) is a call we made through OpenRouter on 2026-11-20.
| Jev | Drex Max | |
|---|---|---|
| Made by | TypeSafe | Nace.AI |
| Model id | ~typesafe/jev-latest | nace-ai/drex-max |
| Weights | Closed | Open RAIL-style for some releases (vendor) |
| List price, input | $0.042 / M tokens (vendor) | $0.06 / M tokens (vendor) |
| Output tokens | Free | Free |
| Context | ~32k shared; 262k opt-in (vendor) | 262k (vendor) |
| Images | Opt-in preview (vendor), not on JevStation | Yes on the API; text only on JevStation |
| How the state is billed | Once | Once (ours) |
| Input tokens, our fit | ≈ 260 + 0.163 × characters + 43 × questions | ≈ 0.163 × characters + 96 × questions |
| Median latency (Banking77) | 367 ms (ours) | 703 ms (ours) |
| Banking77 card intents | 86.5% (ours) | 91.7% (ours) |
| AG News topics | 90% (ours) | 94% (ours) |
| Credits on JevStation | 1 / 3 | 1 / 3 |
Credits are standard / large; an evaluation is large above 8,000 characters of state or five questions.
Accuracy and calibration
We sent both models the same labelled inputs, zero-shot, with one line of description per option. The method is on Decision models compared.
| Task | Jev | Drex Max |
|---|---|---|
| Banking77, 12 card intents (96) | 86.5% | 91.7% |
| AG News, 4 topics (100) | 90% | 94% |
| Banking77: answers at ≥ 0.9 confidence | 78% of inputs, 96% right | 82% of inputs, 100% right |
| Expected calibration error, Banking77 | 0.058 | 0.078 |
| Expected calibration error, AG News | 0.065 | 0.045 |
Drex Max leads on both accuracy tasks. On Banking77 its high-confidence band was cleaner: every answer at 0.9 or above was correct. ECE is mixed — worse on Banking77, better on AG News — so set thresholds on your own labels.
Both models found a single cancel request hidden anywhere in about 20,000 characters of neutral text (Drex Max 0.998 at the very end, control 0.0003).
Tokens, latency and cost
The list price favours Jev ($0.042 against $0.06), but the bill follows the token count. Drex Max counts fewer tokens for the same request and, like Jev, reads the state once. We fitted it on nine requests (state of 100, 2,000 and 8,000 characters, with one, three and five questions):
| Request | Jev tokens | Drex Max tokens | Cheaper at list price |
|---|---|---|---|
| 1 question, short ticket (~300 chars) | ~350 | ~140 | Drex Max |
| 3 questions, 1,500 characters | ~640 | ~530 | close |
| 5 questions, 8,000 characters | ~1,780 | ~1,780 | Jev (lower $/M) |
| 12 rich questions, 24,000 characters | — | 8,927 (ours) | — |
On the AG News sample, mean input tokens were 176 for Drex Max against 412 for Jev, so per thousand calls at list price Drex was about $0.011 against Jev’s $0.017.
Latency favours Jev: Banking77 p50 367 ms against 703 ms for Drex Max, p90 395 ms against 1,033 ms.
When to pick which
- Highest accuracy on our Banking77 / AG News panel among models we host: Drex Max.
- Lowest latency: Jev.
- Lowest list price among one-pass models we host: Perplexity Decider v1.2 at $0.02 / M, then Jev / Decision-1 at $0.042.
- TypeSafe tooling and the original System One product: Jev.
- Open weights under Apache for self-hosting: Clef / Clef-flash.
- Commercial use of a top Nimble checkpoint: not Nimble v4.1 Max (CC BY-NC); use Drex Max or another commercial model instead.
Try both
Run Drex Max above with a free account (50 credits, no card), or use the side-by-side runner to send the same text and questions to Jev and Drex Max at once. The pricing calculator shows the cost of each for your own numbers.
Sources
- Drex Max model, price and context. OpenRouter model page (vendor)
- Nimble licence (non-commercial). Hugging Face Nimble-v4-max (vendor)
- Jev price and context. TypeSafe model docs (vendor)
- Accuracy, calibration, latency and token counts. Our runs through OpenRouter on 2026-11-20 (ours)
JevStation is independent. Jev and Drex Max are names of their respective owners, TypeSafe and Nace.AI; we are not affiliated with either.
Frequently asked questions
- Is Drex Max a drop-in replacement for Jev?
- For the request format, yes. Through OpenRouter’s Decisions / System One endpoint or JevStation it accepts the same state and the same Choice, Score and Noul questions and returns answers in the same shape. Switch models by changing one model field to drex-max.
- Is Drex Max more accurate than Jev?
- In our test on 2026-11-20, yes. Drex Max scored 91.7% on 96 Banking77 card-intent messages against Jev’s 86.5%, and 94% against 90% on 100 AG News articles. Gaps of about five points on samples this size are outside typical noise. Confirm on your own labelled data.
- Is Drex Max better calibrated than Jev?
- On Banking77 every answer at 0.9 confidence or above was right (79 of 79), against 96% for Jev. Expected calibration error was a little worse for Drex (0.078 against 0.058). Treat confidence as a sign, not a guarantee, and check on your own data.
- How does Drex Max compare to Bespoke Nimble?
- Nimble v4.1 Max has posted higher Decision Index scores in some editions, but it is CC BY-NC 4.0 and cannot be used commercially. Drex Max is the commercial path we wire into JevStation. See [Bespoke Nimble vs Drex](/blog/bespoke-nimble-vs-drex).
- What does Drex Max cost on JevStation?
- 1 credit for a standard evaluation and 3 for a large one, the same as Jev. List price on OpenRouter is $0.06 per million input tokens. Because it counts fewer tokens than Jev on the same requests, per-call cost at list price was lower in our AG News sample ($0.011 against $0.017 per thousand).
Related
- Decision modelsJev, Clef, Decider, Luna, Decision-1, Liquid d1 and Drex Max: accuracy, calibration, cost and latency.
- Jev vs DeciderPerplexity Decider against Jev: same contract, measured accuracy, calibration, latency and token counts.
- Jev vs d1-ultraLiquid AI d1-ultra and d1 against Jev: same contract, measured accuracy, calibration, latency and tokens.
- Model pricingMonthly cost of nine decision models for your volume, text length and question count.
Make it part of your pipeline
Sign up for 15 free credits, save your own question sets and call the same evaluation from the API.