On this page
Drex Max (Nace.AI) and Jev (TypeSafe) are decision models: each picks one answer from a list. Drex Max got 94.4% of our 450 test decisions right, and Jev got 86.2%. Both ran on DecideBench v2, 9 November 2026. Drex Max costs $91.00 per million decisions, against $75.04 for Jev. It is also faster: a median answer took 258 ms, against 334 ms for Jev.
If your decisions follow a written policy, such as refunds or release rules, pick Drex Max.
If you already run Jev on short, clear-cut picks it gets right, keep it and test Drex Max on your own data first.
If price per decision matters more than the last few points, look past both: system one model pricing covers the cheaper models.
Both are decision models: each reads some text, a question and a short list of options, and returns one option. Jev, from TypeSafe, made the category popular; what is Jev explains it from scratch. Drex is Nace.AI's family of decision models, and Drex Max is the most accurate one. The full field of 26 models is in the best system one models of 2026.
DecideBench is ours, and so is AI Space, the platform we called Jev through. The test, the data and every answer are public.
Drex Max vs Jev at a glance
Drex Max wins on accuracy and speed; Jev wins on price per token and per decision.
| Drex Max | Jev | |
|---|---|---|
| Maker | Nace.AI | TypeSafe |
| Model we were served | drex-max-v1.1 | typesafe-jev-1.13.0 |
| Right answers (450 decisions) | 94.4% | 86.2% |
| Both halves of a pair right | 89.3% | 72.9% |
| Cost per million decisions | $91.00 | $75.04 |
| Median wait per answer | 258 ms | 334 ms |
| List price, per million input tokens | $0.06 (as of 19 Oct 2026) | $0.042 (as of 25 Sep 2026) |
| Output tokens billed? | No | No |
| API endpoint | POST /v1/systemone | POST /v1/systemone |
| How we called it | Nace.AI's own API | AI Space, our gateway |
"Both halves of a pair right" is the stricter score. Every item in the test has a twin that differs by one small, realistic edit that changes the right answer, such as a file attached to a chat instead of a database table. A model that answers from topic words alone gets one twin wrong. On pairs the gap widens from 8.2 points to 16.4.
What is Drex Max v1.1?
drex-max-v1.1 is the model id Nace.AI's API reported when we asked for drex-max. Nace's Drex Max guide says the original Max id still works and routes to Max v1.1, and the response names the successor. Every Drex Max number in this post is that model.
Drex Max reads text and images, according to the same guide. DecideBench only sends text.
Drex 1.5 is not a separate entry any more. When we asked for drex-v1.5, Nace served drex-max-v1.1, and all 450 answers matched Drex Max's. If you are on Drex 1.5 today, you are already getting Drex Max.
How they did on our tests
Drex Max got more right in all nine kinds of decision; the widest gaps are bug reports, refunds and software releases.
DecideBench v2 is 450 multiple-choice decisions in 225 twin pairs, 50 in each of nine kinds of work. Each item is written the way the decision arrives in production: a JSON payload from an agent, a raw email with headers, a release controller's metrics. Where the right answer depends on a written policy, the question carries the whole policy. Each family has 50 items, so a 2-point gap is one answer.
| Kind of decision | Drex Max | Jev | Gap |
|---|---|---|---|
| Triaging a bug report | 94% | 78% | 16 |
| Ruling on a return under a written policy | 90% | 76% | 14 |
| Gating a staged software release | 74% | 60% | 14 |
| Triaging an inbound email | 98% | 88% | 10 |
| Approving an AI agent's proposed action | 100% | 94% | 6 |
| Checking a claim against its sources | 98% | 94% | 4 |
| Moderating a post | 100% | 96% | 4 |
| Reading a customer's intent | 100% | 96% | 4 |
| Routing an agent to a tool | 96% | 94% | 2 |
| Drex Max | Jev | |
|---|---|---|
| Bug report triage | 94% | 78% |
| Returns | 90% | 76% |
| Release gating | 74% | 60% |
| Email triage | 98% | 88% |
| Agent action review | 100% | 94% |
| Claim checking | 98% | 94% |
| Moderation | 100% | 96% |
| Customer intent | 100% | 96% |
| Agent routing | 96% | 94% |
Source: DecideBench v2, measured 9 November 2026
The gaps follow one pattern: the more a decision depends on applying a rule to a number or a date, the further Jev falls behind.
- Bug reports. Of 4 minor-severity reports, Jev got none right and Drex Max got 3. On reports that were not bugs at all, Jev got 4 of 6 and Drex Max 5.
- Returns. Seven requests arrived outside the return window. Drex Max denied 4 of them; Jev denied 1.
- Releases. These items need a crash-free rate or error ratio checked against an ordered list of rules. Claude Opus 4.8, a large general-purpose model we ran as a yardstick, got all 50 right. Drex Max got 37 and Jev 30. When the right call was to keep the release where it was and wait for more data, Jev chose it 4 times out of 10 and Drex Max 9.
- Email. Of 6 emails that could wait until later in the week, Jev placed 3 correctly and Drex Max 5.
Jev is ahead on a few single answers. It promoted 13 of 15 healthy releases, where Drex Max promoted 11. It graded all 6 degraded-service bugs correctly, and Drex Max got 5. It approved all 7 account-credit refunds, and Drex Max approved 6. These are small counts, and they do not add up to a family Jev wins.
In total Drex Max got 425 of the 450 right and Jev 388. On a million decisions like these, that is about 82,000 more right answers from Drex Max.
Method. We ran both models on the same 450 items on 9 November 2026, sending 4 requests at a time. Each saw one solved example per answer option before every item, from a separate pool of 382 examples. Claude wrote every item; none come from public datasets. Cost is the tokens each API billed, at its dated list price. Latency is the median time per request from one client machine, network included. Older posts on this site quote DecideBench v1, an earlier and easier test; its numbers do not compare with these.
Which does better as the list of options grows?
Drex Max leads at every length of option list, and the gap is widest at four options.
| Options per decision | Items | Drex Max | Jev |
|---|---|---|---|
| 3 | 100 | 99.0% | 94.0% |
| 4 | 150 | 92.7% | 80.0% |
| 5 to 6 | 100 | 87.0% | 82.0% |
| 8 to 10 | 100 | 100.0% | 92.0% |
The long lists are easier than they look. The 8 to 10 option items come from agent routing, customer intent, email triage and bug triage, where the options are distinct teams or tools. The shorter lists hold the policy work: 48 of the 150 four-option items are returns, and 34 of the 100 five-to-six-option items are release decisions. System one model benchmark has the breakdown for every model.
What does a million decisions cost on each?
A million decisions cost $91.00 on Drex Max and $75.04 on Jev, a difference of $15.96.
Neither bills for output, so the whole bill is input tokens. Drex Max's list price is $0.06 per million input tokens, 43% above Jev's $0.042. Drex Max billed fewer tokens, though: about 1,517 per decision to Jev's 1,787. The higher token price outweighs that, so Drex Max ends up 21% dearer per decision.
The token counts are large because of the worked examples. Each decision carried one solved example per option, 5.28 on average, and the inputs are long to begin with. The two APIs count tokens their own way, and our test code sends the examples in a different place for each (see the next section). So 1,517 against 1,787 is what each API billed, not proof that one prompt was longer.
| Volume | Drex Max | Jev |
|---|---|---|
| 100,000 decisions a month | $9.10 | $7.50 |
| 1 million a month | $91.00 | $75.04 |
| 10 million a month | $910.00 | $750.40 |
At 10 million decisions a month, the extra $159.60 buys about 820,000 more right answers. Jev pricing covers how Jev bills in more detail.
Is switching from Jev to Drex Max easy?
Mostly yes: both take a POST to /v1/systemone, and in our test code the same request body and the same answer parser worked against both.
Our test code calls Drex with a subclass of its Jev client. Three things change: the base URL, the API key, and the model id. One thing needed handling. Jev accepted worked examples inside each option's description, while Drex took option descriptions only as plain strings, so we moved the examples into the question's instructions. If you send Jev examples inside the options today, plan for that change.
Nace's Drex Max guide adds two things to know. Max models do not fall back to another model when a request exceeds a limit, and the "latest" alias does not point to a Max model, so name it explicitly.
When to pick each
Pick Drex Max when the decision applies a rule to numbers, dates or severity levels: refunds, release gates, bug triage, inbound email. That is where it gains 10 to 16 points.
Pick Jev when you already run it and your decisions look like the families where the two are close: routing an agent to a tool, moderating posts, reading a customer's intent, checking a claim. There the gap is 2 to 4 points, which is one or two answers in 50. Jev is also 17.5% cheaper per decision on our items.
Before you switch, run your own decisions through both. How to test a classifier walks through doing that with a few hundred labelled examples. If you want a rival not covered here, the best Jev alternatives lists more.
AI Space, our platform, hosts Jev and open decision models on Western infrastructure. We also build agentic workflows that use decision models like these.
What this comparison does not cover
- Jev without worked examples. We planned to run it, but our AI Space account for Jev ran out of credits before that run. Both models here saw examples.
- The gateway hop. We called Jev through AI Space, which adds a hop that Drex Max did not have. Part of the 76 ms latency gap may be that hop; we did not measure how much.
- Sample size. 450 items, 50 per family. Family numbers carry wide error bars; the 37-answer gap across all 450 is the firmer number.
- Who wrote the test. A language model wrote the items and the policies they cite, and three Claude models checked the labels. A careful reader could dispute a few.
- Images. Drex Max reads images; DecideBench does not test that.
- Prices change. Both list prices are dated above. Check Nace's and TypeSafe's pages before you budget.
FAQ
Is Drex Max better than Jev?
On our test, yes. Drex Max got 94.4% of 450 decisions right to Jev's 86.2%, and was ahead in all nine kinds of decision we tested. It was also faster, at 258 ms to Jev's 334 ms. Jev is cheaper, at $75.04 per million decisions to Drex Max's $91.00. On simple routing or moderation the two are within 2 to 4 points.
How much does Drex Max cost?
Nace.AI listed Drex Max at $0.06 per million input tokens on 19 October 2026, with output tokens free. On DecideBench v2, with one worked example per option in every prompt, that came to $91.00 per million decisions. Your cost depends on how long your inputs are and whether you send examples.
Drex Max vs Drex 2.0: which should I use?
Drex Max. Drex 2.0 is another model in Nace.AI's Drex family. On DecideBench v2 Drex Max got 94.4% right to Drex 2.0's 80.9%, at the same $0.06 list price per million input tokens. Per million decisions it cost $91.00 to Drex 2.0's $87.07, with a median wait of 258 ms to 264 ms. The 13.5-point gain costs 4.5% more per decision.
Who makes Drex?
Nace.AI makes Drex. Its website presents Drex as its family of decision models and lists https://console.nace.ai/v1/systemone as the decision endpoint. Drex Max is the most accurate model in the family. Drex 1.5 requests are now served by Drex Max v1.1.
Can I use my Jev code with Drex Max?
Largely. Both answer POST /v1/systemone, and in our test code the same request and response handling worked for both once we changed the URL, key and model id. Drex took option descriptions as plain strings, so worked examples had to move into the question's instructions.
Sources
- XY Space, DecideBench (GitHub repository, v2 results, method and limitations), 9 November 2026
- XY Space, DecideBench leaderboard, Hugging Face, 9 November 2026
- XY Space, DecideBench dataset, Hugging Face, 9 November 2026
- Nace.AI, Drex Max guide, price as of 19 October 2026
- Nace.AI, nace.ai, accessed 9 November 2026
- TypeSafe, Introducing System One Models & Jev, 15 September 2026
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project