On this page
Jev is TypeSafe's decision model: it reads text and picks one answer from a list, such as approve this refund or not. The best Jev alternatives are OpenAI's Decisions API, a paid hosted service, and imajev-4b, which you run yourself. On DecideBench, our public test of 400 decisions such as whether a refund falls inside a 30-day window, Jev scored 98.0%, OpenAI's service 97.2% and imajev-4b 95.0%.
A decision model is a small AI model built for one kind of job: read some text, a question and a short list of options, then pick one option. "Which team gets this support ticket" is a typical question. Jev made the category popular, and what is Jev explains it from scratch. We added OpenAI's service to our results on 8 October 2026.
Want a hosted API that is nearly as accurate as Jev? OpenAI's Decisions API.
Need the model on your own servers? imajev-4b.
Want the cheapest hosted service that gets 9 in 10 right? Drex 1.5.
Want the cheapest model that still gets 4 in 5 right? Jeff Gemma4-E2B.
Already on Cloudflare? Clef, at 94.0%.
Why 19 models? We made 20 decision-model runs. Tev ran twice, once hosted and once on our own graphics card, so it counts once. The three general models and the random baseline in the full table are not in the 19.
- Jev is still the cheapest model above 95%, at $32.26 per million decisions.
- Every price is per million decisions, so hosted and self-hosted models compare directly.
- We ran every test between 28 September and 9 October 2026. DecideBench is ours, and its data is public.
What changed this week?
Three things changed: OpenAI launched its Decisions API, Cloudflare's Clef got cheaper and faster, and Nace.AI open-sourced Drex 1.5.
OpenAI released its Decisions API in public beta on 6 October 2026. It runs one model, GPT-6 Luna, and charges $0.10 per million input tokens with nothing for output. We added it to DecideBench on 8 October. For the product, read what is OpenAI's Decisions API. For the head-to-head, read OpenAI Decisions API vs Jev.
Cloudflare cut Clef-flash's price from $0.09 to $0.038 per million input tokens and made both Clef models faster. We re-ran them on 9 October. Clef's median wait fell from 811 ms to 484 ms. Its accuracy moved from 94.8% to 94.0%, a change of three answers. Clef vs Jev has the detail.
Nace.AI open-sourced Drex 1.5, a model of about 9 billion parameters, on 9 October. It accepts the same request format as Jev, so code written for Jev should work against it. We tested Nace.AI's hosted version.
What are the best Jev alternatives?
The best pick depends on what you are trading away from Jev: a little accuracy for a different vendor, more accuracy for control, or most of it for price. Each pick below comes from the measured numbers.
| If you want | Pick | Right answers | Pairs both right | Cost per million | Median wait |
|---|---|---|---|---|---|
| The closest hosted rival | GPT-6 Luna (OpenAI Decisions API) | 97.2% | 94.5% | $52.82 | 126 ms |
| The best model you run yourself | imajev-4b | 95.0% | 90.5% | $28.57 | 499 ms |
| The cheapest hosted model at 90% or above | Drex 1.5 (Nace.AI) | 90.0% | 80.5% | $23.37 | 147 ms |
| The cheapest model above 80% | Jeff Gemma4-E2B | 80.8% | 64.5% | $16.49 | 76 ms |
| A model on Cloudflare | Clef 27B | 94.0% | 88.0% | $130.85 | 484 ms |
| The model you are replacing | Jev | 98.0% | 96.0% | $32.26 | 639 ms |
"Pairs both right" is the stricter score. Every test item has a twin that differs by one detail that flips the answer. For example, a refund requested on day 30 vs day 31 of a 30-day window. A model that answers from topic words alone gets one twin wrong.
| Right answers | Both halves of a pair right | |
|---|---|---|
| Jeff Gemma4-E2B | 80.8% | 64.5% |
| Drex 1.5 | 90% | 80.5% |
| Tev | 92.8% | 86% |
| Clef 27B | 94% | 88% |
| imajev-4b | 95% | 90.5% |
| GPT-6 Luna | 97.2% | 94.5% |
| Jev | 98% | 96% |
Source: DecideBench v1.1, results measured 28 September to 9 October 2026
Luna, the model behind OpenAI's Decisions API, is the only alternative within a point of Jev. It costs $52.82 per million decisions, against Jev's $32.26. More than half of that bill is worked examples: one solved example for each answer option, sent with every question. Without worked examples it scored 97.0%, and the bill would fall to about $25. Ruling on returns was its weakest job, at 94%.
imajev-4b is a 4-billion-parameter open model. It beats Together AI's Tev on both accuracy and cost, and on our numbers it costs less to run than calling Jev. It got every agent-routing and customer-intent question right.
Drex 1.5 answers fast, at 147 ms. The gap shows in the pairs: 80.5% of pairs both right, against Tev's 86.0%. Moderating posts was its weakest job, at 80%.
Jeff Gemma4-E2B is small and fast, at 76 ms on our card. It struggles with written policies: it got 58% of the returns questions right.
Clef costs about four times as much as Jev and scores four points lower. It makes sense if your stack already runs on Cloudflare's Workers AI. Its smaller sibling, Clef-flash, scores 85.2% at $20.72 per million decisions, which makes it the cheapest hosted entry of all. Its weak spot is checking a claim against evidence, at 66%.
Which Jev alternatives can you run yourself?
Most of the decision models we tested can run on your own hardware. Every entry marked "you" in the full table below is open weights: you can download it and run it on your own graphics card. Clef and Drex 1.5 are open weights too. We tested their hosted versions.
Among the models we ran ourselves, imajev-4b leads at 95.0%. Our own model, Yev, scored 94.2%. Bespoke-Nimble-9B matched Clef's accuracy but cost more than twice as much as imajev-4b, because it is more than twice the size. Tev, JevK5 and Decider-4B follow at 88% to 93%.
The reasons to run one yourself are control and privacy, not price. Yev: an open source Jev alternative covers that case in detail, including how to serve it.
We priced self-hosted models by graphics-card time. Each ran on one NVIDIA L4 graphics card, which rents for about $0.81 an hour, with four requests in flight. A server that batches requests would be cheaper at higher load.
Which is cheapest, and which is fastest?
Jev itself is the cheapest model above 95%. Luna is the fastest hosted model we measured. Below that, the cheapest choice depends on how much accuracy you can give up.
The useful idea here is the efficient frontier. A model is on it when no other model is both cheaper and more accurate. Anything off the frontier is beaten on both counts by something else, so you can ignore it.
The cost frontier, from the top:
| Model | Right answers | Cost per million |
|---|---|---|
| DeepSeek-V4-Flash (general model) | 99.8% | $197.28 |
| GLM-5.3-Flash (general model) | 99.2% | $192.19 |
| Jev | 98.0% | $32.26 |
| imajev-4b | 95.0% | $28.57 |
| Drex 1.5 | 90.0% | $23.37 |
| Clef-flash 9B | 85.2% | $20.72 |
| Jeff Gemma4-E2B | 80.8% | $16.49 |
| Jeff Qwen3.5-0.8B | 71.2% | $9.84 |
Below 70%, two more models sit on the frontier, but they get fewer than 4 in 10 pairs fully right.
The wait times come with two caveats. First, the self-hosted wait times were measured on the graphics card itself, with no network in them. Jeff Qwen3.5-0.8B answers in 43 ms there, but you would add your own network on top. Second, Luna's 126 ms was measured from a cloud data centre, while the other hosted models were timed from one client machine. Luna may well be the fastest, but the comparison is not yet like for like.
Among hosted models timed the same way, Drex 1.5 was fastest at 147 ms, then Tev on Together AI at 197 ms. Jev and Clef were called through AI Space, our model gateway, so their times of 639 ms and 484 ms include it.
Is a general LLM a better alternative?
A general LLM, the large all-purpose kind of model behind chatbots, is worth it only if one or two extra points are worth six to fourteen times the price. Three general models scored 99.2% to 99.8%, against Jev's 98.0%.
| General model | Right answers | Cost per million | Median wait |
|---|---|---|---|
| DeepSeek-V4-Flash | 99.8% | $197.28 | 1,866 ms |
| DeepSeek-V4.1-Flash | 99.2% | $455.09 | 356 ms |
| GLM-5.3-Flash | 99.2% | $192.19 | 665 ms |
All three ran on Together AI, with room to reason before answering. That reasoning is what you pay for. DeepSeek-V4-Flash, the most accurate, also took almost three times as long as Jev.
A common middle path is a cascade: let the decision model answer, and send only its unsure answers to the general model. LLM cascade walks through it, and decision model vs LLM has the full comparison. Jev vs Claude runs the same question against one model family.
How did we test them?
Every model answered the same 400 decisions with the same instructions, and we recorded right answers, cost and wait time together.
The 400 items come in 200 twin pairs, 50 items in each of eight jobs:
- approving an AI agent's proposed action,
- routing an agent to the right tool,
- checking a claim against evidence,
- moderating a post under a written policy,
- ruling on a return under a written policy,
- rating the sentiment of a review,
- reading what a customer wants,
- triaging bug reports.
Each item has between three and six options. Before each item, every model that accepts worked examples gets them, drawn from a separate pool of 297. Chat models answer with a single letter at temperature 0, using Tev's documented instructions. Jev, Tev and Luna were also run without worked examples: Jev scored 98.2%, Luna 97.0% and Tev 90.0%.
Hosted models are priced from the tokens each API billed, at its list price on the day. Self-hosted models are priced by the graphics-card time the whole run took. Every price, source and date is in the dataset's results folder. How to test a classifier explains why the paired design catches models that guess from topic words.
What does this test not tell you?
DecideBench is our own benchmark, and we also make one of the models in it, Yev, which ranks below imajev-4b. Its main limits are a model-written test set, a small sample and models we have not yet run.
A language model, Claude, wrote every test item. We checked the labels by script and by hand, and a careful reader could dispute a few. Jev and Tev miss the same six items, and three of those are arguable.
Each job has only 50 items, so the per-job scores above could swing several points on a new sample and are best read as directional.
Four small models score 35% to 60%. These are encoders, a simpler kind of model, and they often give both twins in a pair the same answer, so they miss the edit that matters.
Nace.AI posted Drex 2.0, a larger model, after our run. We have not tested it.
Hosted wait times depend on where you call from and on the provider's load that day. And the items are public, so a model trained on them later could score higher than it should.
Every model we tested
The table lists all 25 entries: 20 decision-model runs (19 models, since Tev ran twice), three general models and a random baseline. "You" means we ran the open weights on one rented graphics card.
| Model | Who runs it | Right answers | Pairs both right | Cost per million | Median wait |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash (general model) | Together AI | 99.8% | 99.5% | $197.28 | 1,866 ms |
| DeepSeek-V4.1-Flash (general model) | Together AI | 99.2% | 98.5% | $455.09 | 356 ms |
| GLM-5.3-Flash (general model) | Together AI | 99.2% | 98.5% | $192.19 | 665 ms |
| Jev | TypeSafe, via AI Space | 98.0% | 96.0% | $32.26 | 639 ms |
| GPT-6 Luna | OpenAI Decisions API | 97.2% | 94.5% | $52.82 | 126 ms |
| imajev-4b | You | 95.0% | 90.5% | $28.57 | 499 ms |
| yev0-4b | You | 94.2% | 88.5% | $58.10 | 1,058 ms |
| Clef 27B | Cloudflare, via AI Space | 94.0% | 88.0% | $130.85 | 484 ms |
| Bespoke-Nimble-9B | You | 94.0% | 88.0% | $65.30 | 1,116 ms |
| Tev | You | 92.8% | 86.0% | $46.82 | 823 ms |
| Tev | Together AI | 92.8% | 86.0% | $50.20 | 197 ms |
| Drex 1.5 | Nace.AI | 90.0% | 80.5% | $23.37 | 147 ms |
| JevK5 v0.3 | You | 88.8% | 78.5% | $31.74 | 568 ms |
| Decider-4B | You | 88.5% | 78.5% | $29.79 | 525 ms |
| Clef-flash 9B | Cloudflare, via AI Space | 85.2% | 72.0% | $20.72 | 417 ms |
| Jeff Gemma4-E2B | You | 80.8% | 64.5% | $16.49 | 76 ms |
| Kev-4B | You | 78.2% | 65.5% | $26.91 | 437 ms |
| Kev-9B | You | 72.8% | 53.5% | $50.91 | 877 ms |
| Jeff Qwen3.5-0.8B | You | 71.2% | 49.5% | $9.84 | 43 ms |
| Jeff Qwen3.5-2B | You | 70.0% | 48.5% | $12.39 | 52 ms |
| Decider-2B | You | 63.2% | 41.0% | $12.60 | 220 ms |
| Laya typed-decisions 421M | You | 59.8% | 36.5% | $5.45 | 97 ms |
| GLiNER2.5-Decide 340M | You | 57.0% | 30.5% | $14.96 | 255 ms |
| CLM-v0.1-8B | You | 41.0% | 11.0% | $10.46 | 156 ms |
| Julia-1 144M | You | 34.8% | 8.0% | $3.31 | 58 ms |
| Random guess | n/a | 22.2% | 4.5% | $0.00 | 0 ms |
Self-hosted wait times have no network in them. The leaderboard has charts of cost and speed against accuracy, and per-job scores for every row.
FAQ
What is the best alternative to Jev?
For a hosted API, OpenAI's Decisions API, which scored 97.2% on DecideBench against Jev's 98.0%, at $52.82 per million decisions. For a model on your own hardware, imajev-4b, which scored 95.0% at $28.57. No alternative we tested beat Jev on both accuracy and cost.
Is there a free Jev alternative?
Several models are free to download, including imajev-4b, Yev, Tev, Clef and Drex 1.5. You still pay for the graphics card that runs them. On one rented graphics card, imajev-4b came to $28.57 per million decisions, slightly below Jev's API price of $32.26.
What is the cheapest Jev alternative?
Among hosted services, Clef-flash costs least, at $20.72 per million decisions, but scores 85.2%. Drex 1.5 costs $23.37 and scores 90.0%. Among models you run yourself, the Gemma-based Jeff model is the cheapest above 80%, at $16.49. Cheaper models exist, but they score about 71% or lower.
Is OpenAI's Decisions API better than Jev?
Not on accuracy. It scored 97.2% to Jev's 98.0%. On pairs, the gap was 94.5% to 96.0%. It answered faster in our run, at 126 ms, but we measured it from a cloud data centre, so that comparison is not like for like. The API is still in public beta.
Is Tev a good Jev alternative?
Tev is a solid open option at 92.8%, and on Together AI it answers in 197 ms. imajev-4b beats it on accuracy and cost when both run on your own card. Jev vs Tev compares the two in depth.
Sources
- XY Space, DecideBench dataset and results, v1.1, results measured 28 September to 9 October 2026
- XY Space, DecideBench leaderboard, accessed 14 October 2026
- XY Space, DecideBench changelog, 9 October 2026
- OpenAI, Decisions API guide, accessed 14 October 2026
- Unite.AI, OpenAI Releases Decisions API in Public Beta, 6 October 2026
- Cloudflare, Introducing Clef: our open-source decision models, 1 October 2026
- Nace.AI, drex-v1.5 model card, accessed 14 October 2026
- MarkTechPost, Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text, 9 October 2026
- Nace.AI, drex-v2.0 model card, accessed 14 October 2026
- XY Space, yev0-4b model card, 4 October 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