The Jev Playbook · Updated September 29, 2026

Jev cannot write a word. That is the whole point.

TypeSafe AI shipped a model that only ever does one thing: picks an answer off a list you wrote, and tells you how confident it is. No paragraphs, no reasoning, no parsing. About 200 milliseconds and a fraction of a cent. Here is what the creators found when they put it to work, across 34 breakdowns.

TypeSafe AIDecision modelChoice · Score · NullAgent routing34 breakdowns and counting

Read the 60-second version ↓Browse the full library

~200ms

per decision, whatever the schema’s complexity

$0.042 / M

input tokens. Output is free, there is no output

18¢

1,700 emails sorted, scored and spam-checked

39 / 40

head-to-head accuracy, tied with the frontier models

64K

context window. About 6% of a frontier model’s

breakdowns synthesized for this page

The 60-second version

→Jev is an AI model that cannot write a sentence. You give it an input and a list of allowed answers, and it gives back a probability for each one. That limitation is the entire product.
→Every request is one of three shapes: choice, score, or null (a yes/no expressed as a probability). One call can answer several questions about the same input at once.
→Because there are no output tokens, there is no output cost. Input runs about $0.042 per million tokens, and a decision lands in roughly 200 to 400 milliseconds instead of 10 to 30 seconds.
→The move everyone landed on is the same: put Jev in front of the expensive model. It triages the queue, routes the request, or blocks the risky action, and the frontier model only sees what actually needs thinking. On the same 12 support messages, that took Opus 5.5 from 12 calls to 4.
→Read the confidence number, not just the answer. At 90% or higher Jev agreed with the frontier models almost every time in one creator’s tests. Below that, hand the call to a human or a bigger model.
→The limits are real and creators said so out loud: no reasoning, no explanation, a small context window, weak at math and dates. It lost a chess game to GPT-6 Astra in 18 moves, and in TypeSafe’s own four-workflow table it was cheapest and fastest on every row and most accurate on none. This is a traffic cop, not a strategist.

How it actually works

Three shapes. That is the entire API.

You hand Jev an input and a schema of allowed answers. It hands back a probability for each one, in parallel, with no reasoning in between. Every request you will ever make is one of these three.

Choice

Pick exactly one option from a list you define in advance, up to 255 of them.

Sort this email into Sales, Support, Personal, or Spam.

Score

Place the input on a numeric scale you define in advance.

Rate how frustrated this support ticket sounds, 1 to 5.

Null

Answer yes or no, returned as a calibrated probability.

Is this a scam? 91% confident yes.

One call, many questions. The same request can return a category, a 1 to 5 score, and an urgency probability about the same input. You skip training a separate classifier per decision type. In one test, 13 questions in a single call came out over 11 times cheaper and almost 10 times faster than asking them one at a time.

No autoregressive loop. A chat model writes one token at a time and can go wrong on the first one. Jev evaluates every typed question against the state at once.

Trained on calibration, not confidence. RLCD rewards probability estimates that match reality, so an 80% call is supposed to land about 80% of the time.

Worth knowing before you search: one breakdown in this set transcribes the lab as “TypeSpace.” The rest say TypeSafe AI. The model itself is named after Jevons paradox: make a capability cheap enough and people use far more of it, not less. “System One” is TypeSafe’s positioning term, borrowed from Kahneman, for the fast gut-check that sits next to a slower model. ↗ Theo ↗ AI Revolution ↗ Nick Puru

People are asking

The questions everyone’s Googling.

Straight answers, pulled from across all 34 breakdowns.

What is Jev?

Jev is an AI model from TypeSafe AI that never writes a word. You hand it an input and a schema of allowed answers, and it hands back a probability for each option. That is the whole product. It cannot chat, cannot explain itself, and cannot write code. Creators call it a decision model or a classifier rather than a language model.

How is Jev different from ChatGPT or Claude?

A chat model generates an answer one token at a time and streams you text you then have to parse. Jev evaluates every option in parallel and returns typed numbers and categories you can drop straight into code. No reasoning trace, no JSON wrapper, no parsing step. The practical difference creators kept measuring: roughly 200 to 400 milliseconds per decision instead of 10 to 30 seconds.

How fast and how cheap is Jev really?

TypeSafe claims 20x to 200x faster and 40x to 400x cheaper than a comparable LLM call, around 0.4 seconds and $0.0004 per decision. Those are the vendor’s numbers. Creator tests land in the same neighborhood: 185ms on a smart-home command, about 200ms per email across 1,700 emails, 1,000 emails checked against seven rules in 6 seconds for 9 cents, and 250 emails sorted for just over a cent while Claude Opus 5 took about 14 minutes and $1.39 on the same job. At volume, IndyDevDan’s cost table puts a million classification calls at about $20 on Jev against roughly $11,000 on Fable 5.1. Pricing is asymmetric: output is free because there is no output, and input runs about $0.042 per million tokens.

Who makes Jev?

TypeSafe AI, founded by Diogo Almeida, one of the co-inventors of ChatGPT, who spent two years building Jev before it launched on September 15, 2026. The model is named after Jevons paradox: when steam power got cheaper, total coal use went up, not down, because cheaper capability gets used more. That is the bet here. TypeSafe also calls Jev a “System One” model, borrowing Daniel Kahneman’s term for the fast, intuitive half of thinking, which is the role it wants Jev to play next to a slower reasoning model. One early breakdown transcribed the company as "TypeSpace"; the rest of the corpus says TypeSafe.

What are Jev’s three output types?

Choice, score, and null. Choice picks one option from a list you define, up to 255 options. Score rates the input on a scale you define, say 1 to 5. Null is a yes/no answered as a probability, like 91% confident yes. Every Jev request is one of those three shapes, and one API call can answer several of them about the same input at once. It always answers from your list, even when the real answer is not on it, so add a “not stated” option whenever the answer might be missing.

What can Jev actually do?

Anything that is really a fast judgment call on structured input. Creators have demoed inbox triage, lead scoring, routing each request to the cheapest model that can handle it, blocking destructive bash commands before a coding agent runs them, answering questions about a file without loading it into the agent’s context, dispatching edits inside a video editor, driving a browser by voice, stripping ads and cookie banners off web pages, assembling a UI from a component library in under a second, and playing Doom, Subway Surfers and an indie game in real time.

What should you not use Jev for?

Anything that needs reasoning, writing, or a high-stakes judgment call. In a live buy/hold/sell Bitcoin test, Jev underperformed a frontier reasoning model that could cross-reference the news, and Nate Herk’s Bitcoin paper-trading bot on it was down $26 on a $1,000 balance once fees were counted. It lost a blitz chess game to GPT-6 Astra in 18 moves. TypeSafe’s own docs, as Nick Puru reads them, say it is bad at math, reads dates as text, takes your question literally, and gets worse on long, messy documents. It also cannot explain why it decided something, and its context window is small: 64,000 tokens in most breakdowns, 32,000 in Theo’s.

Is Jev more accurate than Claude or GPT?

No, and TypeSafe’s own numbers say so. In TypeSafe’s published table across four business workflows, Jev was the cheapest and fastest option on every row and the most accurate on none. Creator tests match that: a 39 of 40 tie with Claude Haiku, Claude Sonnet 5 and GPT-6 Astra on outfit decisions, 8 of 8 voice-browser commands against 7 of 8 for Claude Haiku, Gemini 3.8 Flash and GPT-5.6 Luna, 76% against 78.3% on a customer-service task, and 29 of 30 inbound leads, one short of the frontier models. It matches them on plain facts sitting in the text and falls behind as the question needs more judgment.

Is Jev calibrated, or just confident?

TypeSafe trained it with RLCD, reinforcement learning for calibrated decisions, which rewards probability estimates that match real-world frequency instead of rewarding answers that sound sure. The claim is that an 80% call actually happens about 80% of the time. That matters, because a confidence number you can trust is what lets you auto-accept high-confidence picks and escalate the rest. Two cautions from later tests. Confidence is not accuracy, so a sure answer can still be wrong and outcomes need checking on their own. And the number is your trust signal: Nick Puru found that at 90% or higher Jev agreed with the frontier models almost every time, and below that a human or a bigger model belonged in the loop.

How do I get access to Jev?

TypeSafe’s console at typesafe.ai started as a waitlist (one creator got in after four or five hours) and now takes signups directly. The Vercel AI Gateway and OpenRouter serve Jev with an API key too, and Zapier added it as a native app and inside its MCP server.

How do you add Jev to Claude Code?

Through the UserPromptSubmit hook, which intercepts a message before the LLM ever sees it. Every prompt hits Jev first; if Jev can resolve it with a direct tool call, it answers instantly and no LLM tokens are spent, and anything it cannot handle falls through to Claude as normal. Setup is a TypeSafe API key plus TypeSafe’s agent setup prompt and skill file, pasted into a fresh session and wired by the agent itself. Two other routes showed up since launch: the free jev-mcp server gives Claude Code or Codex eleven typed judgment tools like verify, screen and gate, and Zapier’s MCP server exposes Jev as a tool next to Gmail and Slack.

Jev vs a chat model, dimensionalized

Faster and cheaper, at exactly one job.

Not vibes. Here is what happened in the tests creators actually ran.

40 identical decisions, four models

JEV413ms per answer, 39 of 40 correct, and zero frontier tokens spent.

THE FIELDClaude Haiku 1,486ms. GPT-6 Astra 3,269ms. Claude Sonnet 5 4,645ms. Same 39 of 40 score, and 74,923 tokens burned to get there.

↗ Julian Goldie

12 support messages, with and without Jev

JEVOpus 5.5 with Jev in front: 19.7 seconds and 4 Opus calls. Jev settled the simple ones itself and only handed over messages that needed a written reply.

OPUS ALONEOpus 5.5 alone: 33.9 seconds and 12 Opus calls, one per message.

↗ Zubair Trabzada

250 real emails, sorted twice

JEVJust over one cent, at about 200ms an email. A prompt injection telling it to run rm -rf got filed as spam.

OPUS 5Claude Opus 5 took about 14 minutes and $1.39 for the same sort.

↗ Creator Magic

Four business tests, one pattern

JEVInbox, contracts, AI slop and leads. Cheapest and fastest every time: 40 emails in under two seconds for under a cent, and an exact match on plain facts like whether a contract auto-renews.

FRONTIERClaude Fable 5.1 and GPT-5.6 Astra took 15 to 30 seconds on the inbox but pulled ahead on urgency, and Jev landed 29 of 30 leads, one short of them. The more judgment a question needs, the wider their lead.

↗ Nick Puru

586 pages of internal links

JEVA whole site’s link map rebuilt in 45.1 seconds for 21 cents. It also refused to link 139 pages where nothing fit.

OPUS 5Claude Opus 5, same job, same clock: 21 pages.

↗ Julian Goldie

One fraud question

JEVFull probability breakdown in about 100 milliseconds, generated in parallel.

AN LLMAbout 8.5 seconds of token-by-token generation, and the answer degrades if the first tokens head the wrong way.

↗ David Ondrej

Vercel’s production swap

JEVReplaced a cheap classifier model in a live workload: 6x faster, and it saturated their eval.

THE OLD ONEThe classifier it replaced was already the cheap option, not a frontier model.

↗ Rob Shocks

Blitz chess, the honest test

JEVBeat Claude Fable 5.1 only on the clock: Fable spent 6 to 15 seconds a move and ran out of time. Jev did not outplay it.

GPT-6 ASTRAGPT-6 Astra did not need the clock. It checkmated Jev in 18 moves.

↗ Matthew Berman

Reach for JEV when…

The answer is one of a small, fixed set you can write down in advance
You are running the same judgment call across hundreds or thousands of rows
Latency is part of the experience: a button click, a keystroke, a driving decision
You want a confidence number you can threshold on, not a paragraph
The decision sits in front of something expensive and only some of it needs the expensive thing
Something hard to undo is about to happen (a bash command, a publish, an outbound send) and you want a cheap check first
You would otherwise be writing a brittle pile of if/else rules

Reach for something else when…

The output is writing: code, copy, an email, a plan
You need to see the reasoning, or defend the decision to a human
The call is high-stakes and worth the extra seconds (the Bitcoin tests and the chess match are the cautionary tales)
The decision needs a lot of reference material: the context window is small (64K, or 32K by Theo’s count)
The answer depends on math or on comparing dates, two weak spots TypeSafe’s own docs admit
The question is genuinely open-ended and the option list cannot be written down

⚠️ The headline 0% hallucination and 0% tool-error figures come from TypeSafe’s own benchmark charts, not an independent test. One creator put a caption over TypeSafe’s benchmark that read “vendor demo, one benchmark, not a 100x result.” The creator-run head-to-heads all point the same way: Jev ties or trails slightly on accuracy while spending a sliver of the time and money. Speed and price are the proven win. Accuracy is a tie at best, not a leap. ↗ David Ondrej ↗ Julian Goldie ↗ Nick Puru ↗ Eric Siu

What the demos proved

The first two weeks, in numbers.

Every one of these was run live on camera. Tap any to watch the source.

1,700 emails

sorted, scored and spam-checked for 18 cents

20,000 requests

of classification for 85 cents, in one creator’s console

9 of 12

prompts never needed the top model. Routing them first cut cost 70%

13,000 → 2,756

tokens for one agent-memory recall, at about $0.003

1,018 papers

sorted into 24 categories for 8 cents, 256ms each

86K tokens

left of a nearly 1M-token Claude Code session after Jev scored every tool call (Theo disputes the idea)

$0.0002

per computer-use step in the TypeSafe Computer Use repo, against about $0.05 with Opus 5

1,191 calls

to score every sentence of a presidential debate, under 5 cents total

7.1 seconds

for a browser agent to book Zurich to London

50 decisions

for a whole AI Town crowd in 0.6 seconds

55 of 500

emails flagged as likely scams in one extra pass

~$7

to play Doom in real time for an hour

Use-case playbooks

Here’s exactly what they built, and how.

The real methods creators demoed, with the actual tools and the video to watch.

The AI traffic cop: put Jev at the front of the queue

↗ Greg Isenberg & Ryan Vogel

EmailContact formsSupport ticketsLead intake

A.Find the expensive incoming queue. Emails, leads, support tickets, quote requests. Anything where a human currently reads everything to find the few that matter.

B.Ask Jev three questions per item. What is it, how important is it, and what should happen next. One call, three answers, about 200ms.

C.Route by confidence, not by category. High confidence goes to a human now. Middle confidence gets automated or drafted by an LLM. Low confidence gets ignored.

D.Start with $5. A starter credit covered two days of heavy team testing, and $10 could plausibly last three months. The cost of finding out is close to zero.

Stop paying a frontier model to sort your mail

↗ Julian Goldie

Claude CodeGPT-6 AstraAgent pipelines

A.Understand where the money goes. Every new message makes an agent re-read the whole conversation. One real follow-up, "do you have the guide on this as well?", cost 186,000 tokens. A bare "how’s it going" cost 222,000.

B.Spot the quiz questions. Inside one sentence like "I’m giving a keynote tomorrow" there are seven separate closed-list decisions: jacket, shirt, trousers, shoes, extras, room setup. None of them need a frontier model.

C.Write the question and the allowed answers. If you can enumerate the options, it belongs to Jev. If you cannot, it stays with the big model.

D.Set a confidence threshold. Auto-accept anything above roughly 60%, flag the rest back to a human or escalate to the bigger model. The calibration is what makes the threshold meaningful.

Wire Jev into Claude Code so simple prompts never cost a token

↗ Kevin Badi

Claude CodeUserPromptSubmit hookTypeSafe API key

A.Use the UserPromptSubmit hook. It intercepts a message before the LLM sees it, which is the right place to split lookups from reasoning. A prompt instruction telling the model to "be efficient" is not.

B.Let the agent wire itself. Three inputs: a TypeSafe API key, their agent setup prompt, and their skill file. Paste all three into a fresh session and tell the agent to build the hook.

C.Color the cheap path. Jev’s answers come back in a different chat color, so you can see at a glance which replies cost nothing. Verify the savings, do not assume them.

D.Test it on a dataset that is not yours. He uploaded a random steak-eating survey mid-video and got a correct cross-tab back in half a second, which proves it generalizes before you trust it on real data.

Give Claude Code a set of judgment tools

↗ AI Edge

Claude CodeCodexjev-mcpshipwithjev.com

A.Install jev-mcp. One free MCP server gives Claude Code or Codex eleven typed tools, among them verify, screen, find, rerank, decide, review and gate. Each answers in 150 to 500ms for a fraction of a cent.

B.Screen what enters context. In the demo, four fact-check verdicts against a page with a hidden injected instruction cost $0.00062 combined, and the injection got caught and blocked.

C.Keep Claude for the plan. Claude or GPT does the planning and the reasoning. Anything that boils down to a yes/no, a score or a single pick goes to Jev.

D.Mine other people’s builds. shipwithjev.com lists 551-plus community Jev projects. Paste it into Claude Code with your own product context and ask where Jev fits, then audit your past sessions for calls Jev could have made.

Guard your coding agent in 10 levels

↗ IndyDevDan

Coding agentsPre-tool hooksJSON payloadsOpen-source repo

A.Start with a smart if-statement. A yes/no prompt-injection check came back at 99%, 82% and 64% confidence. The 99 can auto-block. The 64 goes to a human.

B.Gate the bash tool first. It is the one tool that can do catastrophic damage. A pre-tool hook asks Jev whether each command is destructive and reversible, and it blocked rm -rf and a forced git push even after the agent was talked into trying them.

C.Ask about files without opening them. Does this file validate tokens? Does it hold credentials? Jev answers without the file ever entering the agent’s context, which cut the agent’s token use to about 2K. Across ten files it cost about seven thousandths of a cent.

D.Let the agent write its own questions. Level 10 is a coding agent that decides by itself when to ask Jev, classifying test failures and checking its own fixes. The code is on GitHub as ten-levels-of-jev.

Route every request to the cheapest model that can handle it

↗ Albert Olgaard

Claude Haiku 4.5Claude Opus 5TypeSafe playgroundClaude skill

A.Score difficulty before you pick a model. Jev rates each incoming question, then the easy ones go to Claude Haiku and the hard ones to Opus. Easy traffic came in about 80% cheaper than running everything through Opus.

B.Check the reply on the way out. The same model scores each drafted answer against a checklist: no financial, legal or medical advice, no refund promises, no trash-talking competitors, no leaked system prompt. A request to repeat the system prompt got blocked.

C.Write the criteria, not just the question. "Is cereal a soup?" came back 22% true. Adding one criterion, that soup must be hot and drinkable, dropped it to 6%.

D.Check the bill. Every test in the video, both demos included, cost $0.001. Running both checks on every request barely moved it.

Let Jev run the video editor

↗ Kev Builds Apps

HyperEditFFmpegWhisperRemotionObsidian

A.Make Jev the dispatcher. Every plain-English prompt in HyperEdit’s chat goes to Jev first, and it picks the tool: FFmpeg for audio and video, Whisper for transcripts, Remotion for animation.

B.Keep Claude as the fallback. Claude only gets called when Jev’s three answer types cannot settle the request, so the token bill shows up on the odd request instead of every one.

C.Edit in sentences. "Extract the audio." "Yellow captions." "Remove dead air." The dead-air cut took the timeline from 2:30 to 2:28 by itself.

D.Point it at your asset vault. A connected Obsidian vault let a keyword search pull matching logos and clips straight onto the timeline. HyperEdit is open source, so the setup is the codebase plus a Jev API key.

Gate every business bot: pass, hold, or refuse

↗ Eric Siu

GrokbotRecruitingOutboundContentSaaS spend

A.One classifier layer, every bot. Recruiting, outbound, content and spend all run through the same Jev gate, so tightening it once improves all four.

B.Build it to refuse. The recruiting gate rejects most candidates by design and sends the rest five AI-fluency questions, including how much of their own money they spend on AI tools each month.

C.Screen before you spend. Do not push a full set of sales-call transcripts through an expensive model. Chunk them, tag each chunk cheaply, and only send the hits up.

D.Keep four verbs apart. Soft, ready, send and publish mean four different things. Drafts above a 0.35 ship-ready score publish on their own, the borderline ones get a human, and anything unsure fails closed.

Steal these

16 things to have Jev do for you today.

Real instructions from real demos. Copy, paste, tweak.

“Score every email in my inbox on four axes at once: category, priority, spam likelihood, and how likely I am to reply.”

via Greg IsenbergWatch →

“Score my inbound contact-form leads 0 to 1 so the 98% ones get a reply today and the vague ones wait.”

via Greg IsenbergWatch →

“Add one more question to the same classifier pass: does this email look like a scam?”

via Riley BrownWatch →

“Read this project’s code and docs and propose up to three places where Jev could replace a slow LLM call with a fast, defined decision.”

via Build Great ProductsWatch →

“Add a pre-tool hook that asks Jev whether each bash command is destructive or irreversible, and block it before my agent runs it.”

via IndyDevDanWatch →

“Ask Jev whether this file contains credentials, without loading the file into your context.”

via IndyDevDanWatch →

“Wire Jev into my Claude Code UserPromptSubmit hook so data-lookup prompts never reach the LLM.”

via Kevin BadiWatch →

“Score each incoming question’s difficulty and send the easy ones to Haiku and the hard ones to Opus.”

via Albert OlgaardWatch →

“Go through my agent’s workflow and list every decision that is really a closed-list quiz question.”

via Julian GoldieWatch →

“Auto-accept any Jev pick above 60% confidence and flag everything below it back to me.”

via Julian GoldieWatch →

“Gate my inbox with Jev, then draft replies only for the emails it marks as needing one, and ask me before every draft.”

via Creator MagicWatch →

“Before we draft anything: is this topic already live, what is our closest existing page, and is the new idea meaningfully better?”

via Eric SiuWatch →

“Build a golden set of 100 examples with known answers and benchmark Jev against Opus before we trust it in production.”

via Nate HerkWatch →

“Pick one recurring decision, give Jev the options plus ‘not stated’, and log its confidence for a week before it acts on its own.”

via Nick PuruWatch →

“Build me a Typeform-style application form that re-ranks the candidate into a qualification tier live as they answer.”

via David OndrejWatch →

“Take this support ticket and give me three things at once: the category, how frustrated they sound 1 to 5, and whether it is urgent.”

via Rob ShocksWatch →

Getting access

The waitlist is optional.

Three doors, and all of them are open now.

TypeSafe direct. The console at typesafe.ai started as a waitlist (one creator got in after four or five hours) and now takes signups directly. You get a playground, API keys and an agent setup prompt.

Vercel AI Gateway or OpenRouter. Both serve Jev with an API key. Most of the launch-week demos ran on one of the two.

Zapier. Jev is a native Zapier app, so it can make a call inside any of Zapier’s 9,000-plus integrations. Zapier’s MCP server also hands it to Claude Code as a tool.

Budget for testing: one team’s $5 starter credit covered two days of heavy use, and $10 could plausibly last three months. Another creator’s entire video of tests cost $0.001. ↗ Greg Isenberg ↗ Albert Olgaard ↗ Zubair Trabzada ↗ Matthew Berman ↗ typesafe.ai ↗ Vercel AI Gateway

The map

Where the creators agreed, and where they split.

The honest part you only get by watching all of them. Tap any name to watch.

Routers

“Put it in front of the model you are already paying for.”

Julian Goldie Kevin Badi Riley Brown Albert Olgaard Cole Medin Zubair Trabzada The Next New Thing Jack Roberts

Traffic cops

“Put it in front of your inbound queue.”

Greg Isenberg Riley Brown Creator Magic vogel Nate Herk Eric Siu

Guardrail builders

“Put it in front of anything that can do damage.”

IndyDevDan Albert Olgaard AI Edge Creator Magic

Builders

“The model is not the bottleneck any more. Ship something.”

David Ondrej Greg Isenberg Kev Builds Apps Lukas Margerie Matthew Berman Kevin Badi Build Great Products

Architecture explainers

“Understand why the constraint is the feature.”

Rob Shocks Riley Brown David Ondrej Mark Kashef AI Revolution Zubair Trabzada

Stress testers

“Find the edge before you trust it.”

Theo Matthew Berman Nick Puru Nate Herk Eric Siu

✓ What nearly all of them agreed on

It genuinely cannot write. Every breakdown opens on the same reveal, and every one of them treats the limitation as the point rather than a flaw.
Choice, score, null. Nearly every breakdown describes the same three primitives with the same kinds of examples. A couple count “categorize” as a fourth, but it is the same API.
It belongs in front of something expensive. A frontier model, a human, or a queue. Nobody pitched it as a replacement for any of those.
It adds to the LLM instead of replacing it. IndyDevDan, Cole Medin, AI Edge and Zubair Trabzada all say so outright. Jev makes the call, and the big model still does the writing and the thinking.
The cost is close to noise. 18 cents for 1,700 emails, $5 covering days of team testing, $7 for an hour of Doom. The recurring advice is to just go test it.
The speed changes what is buildable. Sub-second decisions make AI usable inside a click, a keystroke, or a driving frame, which a 10-second reply never was.
Read the confidence, not just the answer. From IndyDevDan’s 64% prompt-injection flag to Nick Puru’s 90% line, the number decides whether a pick runs by itself or goes to a human.
Getting in is easy now. TypeSafe’s console takes signups directly, and Vercel AI Gateway, OpenRouter and Zapier all serve it. The early waitlist took hours, not weeks.

✗ Where they split

How fast, exactly? The numbers run from about 70ms to 500ms depending on who ran it and on what. Albert Olgaard measured 77ms of model time and says most of the wait is server overhead. Even Vercel’s production swap gets quoted as 6x faster in one breakdown and 5 to 18x in others. Treat “about 200ms” as the honest middle and benchmark your own schema.
Whose benchmark? The 0% hallucination and 0% tool-error figures come from TypeSafe’s own charts, and so do the 193.6x faster and 444.6x cheaper headlines on its homepage. Nick Puru points out the 444x is the ceiling, not the average. The creator-run head-to-heads since launch (Julian Goldie, Nick Puru, Creator Magic, Jack Roberts) all show the same shape: a big win on speed and price, and a tie or a small loss on accuracy.
Does it actually cut the bill? Zubair Trabzada went from 12 Opus calls to 4, and The Next New Thing cut cost 70% by routing first. Eric Siu argues the other side in his classifier-layer video: Jev is one more model call, so spend goes up slightly, and the payoff is speed and consistency. It saves money when it stops the big model from being called at all.
Can it judge or compact an agent’s work? Mark Kashef pitches it for fact-checks and as a cheaper check on an LLM’s output, and a plugin in Julian Goldie’s course cut a Claude Code session from nearly 1M tokens to 86,000 by scoring every tool call. Theo says both uses misread the model: it never sees tool outputs, reasoning traces or the full history, so it cannot tell what is safe to drop.
Is a classifier really new? Rob Shocks and Riley Brown both address the obvious objection that LLMs already do structured output. Their answer is that generating text, formatting it as JSON, then parsing it back pays for three steps where Jev pays for one. Cole Medin, Mark Kashef and Nick Puru add a sharper one: old classifiers had to be trained per task, and Jev takes plain-English options with no labeled data.
How far up the stack does it go? David Ondrej builds a whole product on it. Greg Isenberg’s own Bitcoin test says stop at routing, and he says so on camera. Nate Herk’s paper-trading bot agrees: down $26 on $1,000 once fees were counted.
“Built with Jev”? Zubair Trabzada calls out demos claiming apps were built with Jev, including a Tesla self-driving rebuild, because Jev cannot write code. Matthew Berman shows that Tesla demo anyway. Both are right: a developer built the simulator, and Jev only makes the driving calls inside it.
Even the basics. One breakdown transcribes the lab as “TypeSpace,” the rest as “TypeSafe AI.” Most of the corpus puts the context window at 64K tokens, and Theo says 32K. Plan for the smaller number until TypeSafe’s docs settle it.

The 8 moves that pay

Audit your agent for quiz questions

Before you wire anything, list every decision your pipeline already makes where the answer comes from a list you could write down. That list is your Jev surface area, and it is usually bigger than you think.

Put the cheap model in front, not behind

The single most-repeated move in the corpus. Jev triages, routes, or answers the lookup. The frontier model only wakes up for what actually needs thinking.

Threshold on confidence

A calibrated probability is only useful if you act on it. Pick a number, auto-accept above it, escalate below it. That one rule turns a classifier into an actual system.

Word the question, not the prompt

The quality of a score depends on how precisely you define the scale and the yes/no anchors. There is no prompt engineering to hide behind here, which is a feature.

Gate anything you cannot undo

A bash command, a publish, an outbound send. Put a cheap yes/no in front of every irreversible action, and keep “ready” and “send” as separate steps. IndyDevDan and Eric Siu both built their systems around this.

Test it on your own data first

Build a small golden set with known right answers, run Jev and your current model on it, and watch the confidence scores for a week before letting it act alone. Nate Herk and Nick Puru both prescribe exactly this.

Make the savings visible

Color the cheap path, log the token count, count the calls. Otherwise you are assuming the savings instead of verifying them, and the whole reason to do this is the bill.

Know where to stop

No reasoning, no explanation, 64K of context. The Bitcoin test is the line: routing decisions yes, high-stakes judgment calls no. Nobody in the corpus argues otherwise.

What’s changed

This page is alive. Here’s the update log.

Every new Jev video we break down gets added to this page automatically. Every couple of weeks we go through the new ones and rewrite the guide itself around what they found. Here’s what changed and when.

September 29, 2026

Second edition. 23 more Jev breakdowns have landed in the library since the first edition, so the guide got rewritten around them. The biggest change: there are now enough creator-run tests to say plainly where Jev wins (speed and price) and where it does not (accuracy on judgment calls).

Five new head-to-heads, including Opus 5.5 with and without Jev, 250 emails against Opus 5, and the chess game Jev lost
A new question answered straight: is Jev more accurate than Claude or GPT?
Five new playbooks: guarding a coding agent, judgment tools for Claude Code, routing to the cheapest model, running a video editor, and gating every business bot
Eight new things to have Jev do, and new demo numbers from the first two weeks
Two new groups on the stance map, guardrail builders and stress testers, with every new creator placed
New splits: whether Jev really cuts the bill, whether it should judge or compact an agent’s work, and how big its context window is
Access updated: signups are open, and Zapier serves Jev now too

↗ IndyDevDan: 10 Levels of Jev For Agentic Engineers ↗ The Next New Thing: 9 things you’ll actually do with Jev ↗ Zubair Trabzada: Opus 5.5 + Jev Might Be the Ultimate AI Combo ↗ Eric Siu: 7 Ways I Use Jev + Grokbot to Run My Business ↗ AI Edge: 7 Jev Repos That 10x Claude Code (Free)↗ Matthew Berman: 8 Jev Use Cases That Feel Like Cheating ↗ Nick Puru: Jev, the AI Model That Only Picks Answers, Never Writes Them ↗ Kev Builds Apps: Jev AI Inside HyperEdit, the Open-Source AI Video Editor ↗ Zubair Trabzada: The Rush Hour Traffic Sim That Shows What Jev Actually Does ↗ Cole Medin: Jev’s System 1 Decisions in His Game, PR Triage and LLM Router ↗ Build Great Products: Jev Just Changed Software Forever ↗ Creator Magic: Jev Sorted 250 Emails for 1 Cent ↗ Julian Goldie: Jev AI Full Course ↗ Eric Siu: Jev, a Classifier Layer for Your Existing AI Agent Stack ↗ Jack Roberts: Jev Isn’t a Chatbot. It’s a Decision Engine That Costs a Penny ↗ Nate Herk: Jev Tested on 12 Use Cases ↗ Albert Olgaard: Jev, a Fast, Cheap Classifier for Routing and Guardrails ↗ Kevin Badi: Master 95% of Jev AI in 7 Minutes ↗ Eric Siu: 5 Real Marketing Use Cases for Jev ↗ Lukas Margerie: A Pixel Painter, a YouTube Classifier and Voice-Controlled Subway Surfers ↗ Mark Kashef: Jev Explained, a Classifier for Fact-Checks, LLM Judging and Routing ↗ Matthew Berman: Jev Is a Decision Engine, Not a Thinker (the Chess Match)↗ vogel: Jev Classifies 1,500 Real Inbox Emails in Real Time

September 20, 2026

First edition. Jev launched days ago and the first creator breakdowns are in, so this page went live built from the launch-week corpus.

The questions people are Googling, answered straight, with the number and the creator attached to each one
The three primitives explained: choice, score, and null
Four head-to-heads creators actually ran, including the only independent one in the set
Five use-case playbooks, from inbox triage to wiring it into Claude Code
12 things to have Jev do for you today, copied from real demos
The honest limits section: no reasoning, 64K context, and the live trading test it lost

↗ Rob Shocks — Jev: TypeSpace’s Decision-Only AI Model Explained ↗ Greg Isenberg — Jev: The Classifier AI That Scored 1,700 Emails for 18 Cents ↗ Riley Brown — JEV: How It Works and What You Can Build ↗ David Ondrej — Build Anything With Jev, TypeSafe AI’s Probability-Only Model ↗ Kevin Badi — How to Add Jev AI to Claude Code (Step by Step)↗ Julian Goldie — Jev AI: How to Stop Wasting Claude and GPT-6 Astra Tokens on Tiny Decisions

The library

All 34 breakdowns, and counting.