Jev vs an LLM
An LLM writes; Jev decides. Use Jev when the answer is one of a set you can write down in advance and the same decision happens many times or under a deadline. Use an LLM for everything that has to be composed. Most builds in this directory do both, and the numbers below are the ones their authors published.
Updated 22 Sept 2026 · by Made with Jev
In short
- Same accuracy as Sonnet 5 on TypeSafe’s workflow evals, 67.8%, at 1/290th of the cost and 1/195th of the time.
- 100,000 posts for $0.67 in 20.4 s, against 214 posts for $0.98 on Claude Opus 5 in the same 20 seconds.
- Jev cannot write, cannot take images, and a single choice holds at most 255 options.
- The larger model still catches what Jev misses. One editorial test found six of seven planted defects with Jev, seven of seven with a frontier model.
The difference, line by line
Everything below follows from one design decision: Jev emits a distribution over answers you defined instead of decoding tokens.
| Jev | A general-purpose LLM | |
|---|---|---|
| What comes back | One of your options, probabilities, a confidence | Text, of whatever length it decides |
| Time to answer | 70 to 500 ms | Seconds to minutes, growing with the output |
| Cost per case, TypeSafe evals | $0.0004 | $0.0195 to $0.1761 |
| Many questions at once | Answered in parallel, ~one question’s cost | Longer prompt, longer output, more money |
| Can it write? | No | Yes, that is the job |
| Input | Text only | Text, images, audio, video, depending on the model |
| Answer space | Closed: at most 255 options in one choice | Open |
| How wrong looks | A wrong option, with a low confidence you can act on | A confident paragraph that may be invented |
Eval figures from TypeSafe’s workflow evals, averaged over four workflows. The reference answers there come from two large reasoning models, not from people, so 67.8% means Jev agreed with them 67.8% of the time. Full table on what is Jev.
Where someone ran the job both ways
These are not benchmarks. They are people who had an LLM doing a job, swapped in Jev, and published both sides.
| The job | With a large model | With Jev |
|---|---|---|
| 100,000 viral posts, 14 questions each | 214 posts, $0.98, Claude Opus 5 | 100,000 posts, 20.4 s, $0.67 |
| 1,709 editorial judgments | 8.83 s per passage, Fable 5.1 | 0.35 s per passage, under a cent in total |
| A 26-sheet construction plan set | Their GPT-4.1 production pipeline | 2.9 s, $0.0052, 17–21× cheaper, 5× faster |
| Playing Slay the Spire 2 | GPT-6 Astra: capable, slow | 0.7 s a move |
| A 30-way labelling task | OpenRouter’s popular models | >5× the next fastest; slowest requests beat their medians |
Movez
@0xMovez
I just built a Jev X Viral Post Analyser. 100,000 viral X posts. 20.4 seconds. $0.67. Claude Opus 5, same corpus, same clock, got through 214 posts and spent $0.98. per post that is ~680x cheaper the full Opus pass would have run $458. viral analysis is the perfect Jev job. • it is not writing, it is 14 yes/no calls per post: > does the hook open a loop, > is there a number in the first line, > is the proof real or claimed. classification, not prose. • what it found: 1,220 posts broke into the top 1%. baseline 1.22%. > superlative claim - 2.34% viral. 1.92x baseline > contrarian take - 1.59%. 1.31x > launch / tool drop - 1.46%. 1.19x and numbered lists, the thing everyone writes: 0.55%. below baseline. the most used hook is the least viral one. full stop. • what you are watching: left is the post under analysis, right is Jev answering 14 typed questions about it, each with a confidence score. the run stops at 20.4s because that is when Jev finished all 100k. pulled the corpus through a few X APIs, one parallel pass into Jev. should I drop it to public? Read my latest article on Jev Engineering below and turn your ideas into reality.
Dan Shipper
@danshipper
we almost never test new foundation models but we've been testing this for ~a week @every and it's pretty wild. the kind of things that will be obviously indispensible in 6-12 months it doesn't produce words as output, it produces probabilities. so it can efficiently act as a judge in cases where you'd need a Fable-level model—but in our testing was 25x faster and 600x lower priced excellent vibe check by @hammer_mt on @every: https://every.to/also-true-for-humans/mini-vibe-check-typesafe-s-jev-judged-everything-i-ve-written-in-0-7-seconds?utm_cta_source=home_main_a_3
ArticleContent and growth
PickEvery’s editorial vibe check
- Judgments
- 1,709
- Total cost
- <$0.01
- Median per passage
- 0.35 s
Trinay Hari
@hari_trinay
Built a construction plan-set classifier with Jev. Proq turns civil and building plan sets into bills of materials using an LLM pipeline we built on GPT-4.1. Jev classified an entire 26-sheet plan set in 2.9 seconds for $0.0052. It matched GPT-4.1 and GPT-6 Astra on 100% of sheet-level classifications while running 17–21x cheaper and 5x faster than our production pipeline.
paulwei
@coolish
昨天出的 typesafe.ai Jev 模型 适合高频决策，那岂不又是适合打游戏？ 之前用 GPT-6 Astra 代打杀戮尖塔2 能力强但速度慢。 我刚实测用 Jev 打，行动思考只需 0.7秒， 画面我都没看清它就操作完了。 这超人类游戏速度， 着实又让我震惊瘫坐了😅 语言很难描述这感觉，看视频⬇️
OpenRouter
@OpenRouter
1/ Jev, a decision model by @typesafeai, sparked a burst of projects and discussion. We tested it using Ori Eval against popular LLMs on OpenRouter at judging. Jev was >5x faster than the next fastest model, and even its slowest requests beat every other model's median.
Hassan
@nutlope
Jev + Kimi K3 for fraud detection! TLDR: Jev classified 100 emails in 1.42 seconds, then I routed the uncertain cases to Kimi K3. The full pipeline got 96/100 correct for only ~$0.07. Video is not sped up, check out the live run! Here was my process: I gave Jev 100 emails to classify (a mix of 50 legit & 50 fraudelent emails). It classified all of them in 1.42 seconds. An underrated feature about Jev is it will give you the confidence score for a classification, so I routed any prediction under 95% confidence to Kimi K3 to be fully sure. 31 emails fell below that threshold. After routing those to Kimi K3, the combined pipeline reached 96% accuracy. The full run took 16 seconds & ~$0.07 in inference costs: - $0.068 from Kimi K3 on @togethercompute - $0.003 (1/3 of a cent) from Jev on @typesafeai. I think this is a really interesting pattern: use a fast specialized model like Jev for the narrow task, then route the uncertain cases to a larger LLM. I feel like this kind of approach could be a game changer for use cases like fraud or anything realtime. You can use the speed & low cost of Jev while having a larger LLM as a fallback to ensure high accuracy.
XTriage and routing
PickFraud detection with Jev and Kimi K3
- Emails
- 100 in 1.42 s
- Correct
- 96/100
- Cost
- ~$0.07
Where the large model still won
The most honest test in the directory is Mike Taylor’s. He planted seven defects across his own writing and ran both. Jev found six of the seven for under a cent. The seventh needed the larger model. That is the trade in one sentence: you are buying almost all of the judgment for almost none of the money, and the part you give up is the part that needed reading between the lines.
Three more limits worth knowing before you swap anything. A choice holds at most 255 options, so high-cardinality problems need a two-pass shortlist. The input is text only, so anything visual needs OCR or a segmentation model in front — which is exactly what the computer-use builds here do. And people have published negative results too: one builder found Jev worse than his existing setup as a reranker and said so.
Will Kelly’s adversarial evaluation is the one to read before you trust any of this: 28 predictions written down before any data, then nine experiments against them, 123,805 requests and $12.69, with what failed published alongside what held.
The answer most builds land on: both
Almost nothing here is Jev instead of an LLM. It is an LLM with its repetitive decisions taken out. A router asks Jev which model should handle a request before paying for any of them. A compaction plugin asks Jev which parts of a Claude session still matter, instead of paying a large model to summarize them. A fraud pipeline lets Jev clear 100 emails in 1.42 seconds and sends only the unsure ones to Kimi K3, scoring 96 out of 100 for about seven cents.
Duncan
@ephraimduncan
Built a model router with Jev by @typesafeai. Jev decides what model fits your request best and the request is sent to that model.
XTriage and routing
A model router on Jev
tamara
@tamarajtran
found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by scoring every tool call and dropping what’s irrelevant
XTools and apps
PickInstant compaction for Claude
Hassan
@nutlope
I used Jev to classify 1,018 AI research papers. The result: $0.08 total cost and 256ms median end-to-end latency per paper. The pipeline was: 1. Summarize each paper with DeepSeek V4 Flash 2. Send the title + summary + 24 possible topics to Jev 3. Use Jev to classify each paper 4. Visualize everything on http://1kpapers.com The summaries cost $3.99 on @togethercompute. The classifications cost $0.08 on @typesafeai. So for just over $4 of inference, I ended up with a pretty useful way to explore the top AI research papers from the past year. I think this is where things are heading: different models for different parts of the workflow, instead of using one model for everything. I’m running evals on the Jev classifications before replacing the current ones, but the site is already live: http://1kpapers.com
CJ (Coding Garden)
@CodingGarden
I built a chat bot with jev, no LLM at all! Responses are instant, no hallucinations. I hooked it up to web search, wikipedia, weather, todoist and home assistant. Jev decides what tool to call and what args to use based on the prompt. Instant answers cite sources as well!
XAgents and browsers
A chat bot with no LLM
The split has a name on this site: Jev Engineering. An LLM writes, Jev decides, code acts.
How to decide, in four questions
- Can you write the acceptable answers down before the call? If not, it is an LLM job.
- Does the decision repeat, or does it have a deadline measured in milliseconds? If neither, the cost difference will not pay for the rewrite.
- Is there something sensible to do with an unsure answer? If there is no escalation path, you lose the main safety feature.
- Is the input text, or can you cheaply turn it into text? Every build here that reads a screen, an image or a PDF puts OCR or a segmentation model in front of Jev first.
And if the answer points at a decision model but not at an API, there are now open ones that do the same job on your own machine: the open alternatives.
If the answers point at Jev, the cost side is on the pricing page and the first call is on how to use Jev.
Common questions
- What is the difference between Jev and an LLM?
- An LLM generates text one token at a time; Jev returns one of the answers you defined, with a probability for each and a confidence score, in a single pass. That is why it answers in 70 to 500 milliseconds and why it cannot write anything.
- Is Jev better than GPT or Claude?
- Not at anything they are for. On TypeSafe's own workflow evals Jev matches Sonnet 5's accuracy at 67.8% while costing $0.0004 a case against $0.1174 and taking 0.4 seconds against 78.1. On writing, reasoning and open-ended work it does not compete, because it does not do them.
- Can Jev replace ChatGPT?
- No. It replaces the classify, route, score and approve calls inside software you already built. A chat product still needs a model that writes.
- When should I use an LLM instead?
- When the set of possible answers is not known in advance, when the output has to be prose or code, when the input is an image or audio, or when the job is one hard judgment rather than a thousand repeated ones.
- Jev vs Laya?
- Laya is not an LLM, it is an open decision model that does the same job as Jev: 421M parameters, running locally in about a gigabyte, with its author reporting 86.5 decisions a second at a P50 of about 9 milliseconds. That comparison belongs on the open-source page, not this one; nobody has yet published a head-to-head on accuracy that someone else could re-run.
Made with Jev is independent and not affiliated with TypeSafe AI. Every figure on this page is the one its author published, linked to where it can be checked.