Big Idea

The argument in one line.

A model that only selects from a supplied list of answers, and never generates new text, can absorb most of an AI agent's tiny repetitive decisions at a fraction of the token cost and time of a frontier model like Claude or GPT-6 Astra.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're running an AI agent (Claude Code, GPT-6 Astra, or similar) that burns through your token plan on small classification or yes/no decisions.
You've noticed a single short message costing tens or hundreds of thousands of tokens because the whole chat history gets re-read every time.
You build agent workflows with repeated small routing steps: email triage, outfit or content picking, browser automation, sorting.

SKIP IF…

You're looking for a model that can write or generate new content — Jev can only select from options you give it, it never produces original text.
Your workflow has no repeatable multiple-choice decisions, only open-ended reasoning, writing, or planning tasks.

TL;DR

The full version, fast.

Every message you send an AI agent forces it to re-read the entire conversation from the start, so short replies can quietly cost hundreds of thousands of tokens. Most of that reading pays for tiny multiple-choice decisions (which shoe, is this email urgent) that don't need a frontier model's judgment. Jev, a decision-only model from TypeSafe AI, answers those questions from a fixed list of options with a confidence score, never writing a word, at about $0.042 per million tokens versus $0.20-$10 for Claude or GPT-6 Astra, and roughly 3-10x faster. The fix is a three-step habit: spot the quiz-like decisions your agent already makes, write out the question and the allowed answers, and route just that step to Jev while the frontier model keeps doing the real writing and planning.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:47

01 · Stop Wasting Tokens

Cold open: big AI models are doing hundreds of tiny jobs they don't need to, and every one eats tokens. Jev AI is introduced as a tiny model that answers those jobs in under half a second.

00:47 – 01:08

02 · Tokens Explained Fast

Defines a token as the unit Claude/GPT-6 Astra use to measure reading; plans cap total tokens, and hitting the limit stops the agent.

01:08 – 01:58

03 · The Hidden Chat Tax

Every message re-reads the entire chat from the top. Real examples from Joe's own agent chats: 124 tokens for a short add-on, 186,000 for a guide question, 222,000 for 'how's it going'.

01:58 – 02:25

04 · Why Big Models Waste

Frontier models are built for planning, writing, and fixing hard problems, not micro-decisions; handing them a tiny choice is like paying a professor to sort the mail.

02:25 – 03:43

05 · Meet Jev AI

Jev, released September 15, 2026 by TypeSafe AI, never writes: given a question and a list of allowed answers, it picks one and returns a confidence score, so it can't hallucinate an answer outside the list.

03:43 – 04:57

06 · Outfit Picker Race

One sentence ('I'm giving a keynote tomorrow') becomes seven tiny wardrobe choices. Jev, Claude Haiku, GPT-6 Astra, and Claude Sonnet 5 race on 40 of these: Jev answers in 413ms and matches the top score at 39/40.

04:57 – 06:20

07 · Real Time Choice Systems

The same pick-from-a-list approach drives a hands-free voice-browser control demo (Moritz Krem's open-source project) and an email-triage demo that sorts 60 emails in 25 seconds using four fixed actions.

06:20 – 07:57

08 · Three Step Setup

The implementation method: spot the quiz-like tiny choices your agent makes, write out the question and the allowed answers, then send just that step to Jev (available via OpenRouter) while the frontier model keeps the real work.

07:57 – 08:20

09 · Confidence Thresholds

Jev's confidence score lets you set a cutoff, e.g. above 60% the agent proceeds automatically, below it the decision gets flagged back to a human or a bigger model.

08:20 – 09:14

10 · Costs Speed and Wrap Up

Cost comparison table: frontier input tokens run $0.20-$10 per million versus $0.042 for Jev, with output effectively free since Jev never generates new text; framed as one of the year's bigger model releases.

09:14 – 09:39

11 · Final Call to Action

Closing pitch for the AI Profit Lab / 'AI Profit Boardroom' community and its Jev AI training.

Atomic Insights

Lines worth screenshotting.

Every new message forces an AI agent to re-read the entire conversation from the first line, not just the newest input, so token cost balloons as a chat grows.
One real follow-up message, 'do you have the guide on this as well?', cost 186,000 tokens because the full chat history came along for the ride.
A one-line reply like 'how's it going' burned 222,000 tokens in a single answer once the whole conversation was re-read to produce it.
Jev never generates text: it's given a question, a list of allowed answers, and returns one pick plus a confidence score.
Because Jev can only choose from the options it's handed, it can't invent an answer that doesn't exist in the app, unlike a model that writes free text.
On 40 identical outfit-choice requests, Jev answered in 413 milliseconds versus 1,486ms for Claude Haiku, 3,269ms for GPT-6 Astra, and 4,645ms for Claude Sonnet 5.
Jev matched the group's top accuracy at 39 out of 40 correct picks while cutting the big models' token use on those decisions from 74,923 to zero.
One sentence like 'I'm giving a keynote tomorrow' can hide seven separate tiny decisions: jacket, shirt, trousers, shoes, extras, and room setup.
Frontier input tokens run $0.20 to $10 per million depending on the model, versus about $0.042 per million for Jev, roughly a 99% reduction on offloaded decisions.
Jev has effectively no output cost because it never generates new text, while a normal LLM's output tokens typically cost about five times more than its input tokens.
A confidence threshold (for example 60%) lets an agent auto-accept Jev's high-confidence picks and automatically flag the low-confidence ones back to a human or a bigger model.
An email-triage demo sorted 60 emails in 25 seconds by picking one of four fixed actions per message: reply, research, wait, or flag.

Takeaway

A tiny AI that only picks can replace a frontier model on tiny decisions

TOKEN ECONOMY

Every small yes/no or which-option decision your agent makes gets billed at full chat-history rereading cost, and a decision-only model that just picks from a list can absorb almost all of that for about 1% of the price.

02Tokens Explained Fast

A token is the unit Claude and GPT-6 Astra use to measure how much text they read, and your plan caps total tokens read per period.
The more text a model has to read, the slower it answers, and hitting the token limit stops your agent cold.

03The Hidden Chat Tax

Every new message forces the AI to re-read the entire conversation from the very first line, not just your latest input.
By message 50, answering one short line means re-reading the equivalent of a full book of accumulated chat history.
A single one-line reply like 'how's it going' burned 222,000 tokens once the whole conversation was re-read to answer it.

04Why Big Models Waste

Frontier models like Claude and GPT-6 Astra are built for planning, writing, and fixing hard problems, not micro-decisions.
Handing a tiny yes/no or which-option choice to a frontier model is like paying a professor to sort your mail.

05Meet Jev AI

Jev, released September 15, 2026 by TypeSafe AI, never writes a word: given a question, a list of allowed answers, and it picks one with a confidence score.
Because it can only pick from the options supplied, Jev can't hallucinate an answer that doesn't exist in your app, unlike a model that writes free text.

06Outfit Picker Race

One sentence like 'I'm giving a keynote tomorrow' hides seven separate tiny decisions: jacket, shirt, trousers, shoes, extras, and room setup.
On 40 identical outfit-choice requests, Jev answered in 413 milliseconds versus 1,486ms for Claude Haiku, 3,269ms for GPT-6 Astra, and 4,645ms for Claude Sonnet 5.
Jev matched the group's top accuracy at 39 out of 40 correct picks while cutting the frontier models' token use on those decisions to zero.

07Real Time Choice Systems

The same pick-from-a-list approach powers hands-free voice-browser control, where voice commands map to a fixed menu of actions instead of a model reasoning through each step.
An email-triage demo sorted 60 emails in 25 seconds by picking one of four fixed actions per message: reply, research, wait, or flag.

08Three Step Setup

Step 1 is auditing your agent's day for quiz-like moments, decisions that could be answered from a short fixed list rather than open-ended reasoning.
Step 2 is writing out the exact question and the full set of allowed answers so Jev has a closed list to choose from.
Step 3 is routing just that one decision to Jev via its API, available through OpenRouter, while the frontier model keeps handling the real writing and planning work.

09Confidence Thresholds

Jev returns a confidence score with every pick, so you can set a threshold like 60%: above it the agent proceeds automatically, below it the decision gets flagged back to a human or a bigger model.

10Costs Speed and Wrap Up

Frontier input tokens run $0.20 to $10 per million tokens depending on the model, versus about $0.042 per million for Jev, roughly a 99% reduction on the offloaded decisions.
Jev has effectively no output cost because it never generates new text, while output tokens on a normal LLM typically cost about five times more than input tokens.

Glossary

Terms worth knowing.

Token: The unit Claude and GPT-6 Astra use to measure how much text they read. Usage plans cap total tokens read over a period, and hitting that cap stops the agent from responding.
Jev: A decision-only AI model from TypeSafe AI, released September 15, 2026, that picks one answer from a fixed list you supply and returns a confidence score, without generating any new text.
Confidence score: The percentage Jev attaches to its chosen answer, used to decide whether to auto-accept the pick or escalate the decision to a human or a larger model.
Tiny choice: A small, closed-list decision inside an AI agent's workflow, such as which email folder or which product option, that doesn't require a frontier model's open-ended reasoning to answer.

Resources

Things they pointed at.

02:25toolJev AI (TypeSafe AI)

07:16toolOpenRouter

05:25toolMoritz Krem's open-source voice-browser control system

06:03productAI Profit Lab ("AI Profit Boardroom") community ↗

Quotables

Lines you could clip.

02:02

“How's it going? It costs 222,000 tokens in one answer because the whole chat came along for the ride.”

a concrete, shocking number that makes the abstract 'wasted tokens' problem visceral→ TikTok hook↗ Tweet quote

01:37

“That's like paying a professor to sort your mail.”

tight metaphor that reframes the whole video's premise in one line→ IG reel cold open↗ Tweet quote

04:16

“39 out of 40, right calls, all four models, same score. 74,923 tokens down to zero.”

the single stat that proves the pitch with a before/after number→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

jev ai how to stop wasting claude and gb6 astra tokens so here's how to reduce your claude and gb6 astra token usage with jev ai so right now your big ai models are doing hundreds of tiny jobs they don't need to do and everyone eats your tokens now jev ai is a tiny ai that takes those jobs answers in under a half second and tells you how sure it is now i actually raced it against for example claude and jeep six and four real jobs and in one race the tiny model got the most answers right you'll see all four races the three easy steps set it up and two bonus moves for your everyday chat so stick with me because one of those bonus moves made my ai agent forget its own rules and the fix was one number let's start simple so what is a token.

A token is how Claude and GP6 Astra measure reading. Your plan only lets the AI read so many tokens. The more it reads, the slower it answers.

And when you hit the limit, your agent stops. So if your AI reads less, it runs faster and it runs longer. Today is about making it read less.

Now, most people don't know this part, but every time you send a message, the AI reads a whole chat again from the top. Your first message, it reads. one page by message 50 it reads a whole book just to answer one line every file opened is still in there so in every test it's going to read that over and over and over again now i actually checked my own chats with my ai agents and these are like real messages are typed so for example i actually typed can you add this one too and it read 124 124 tokens I also asked it, you know, do you have the guide on this as well?

And that took up 186 ,000 tokens. And here's my favorite. How's it going?

It costs 222 ,000 tokens in one answer because the whole chat came along for the ride. Now, here's the real waste. Clone Gypsies are...

like professors. They're great at thinking, planning, writing, fixing tricky things. But your agents also hand them hundreds of tiny choices.

For example, is this email urgent? Which shoes should these two pages link? And that's like paying a professor to sort your mail.

And every one of these tiny choices uses up more tokens. Now, what is Jeff? Well, basically, most AI Models write.

Jeff doesn't write a single word. You can just give it three things. What's going on?

A question, a list of answers. It's allowed to pick from. It picks one and it tells you how sure it is, like 90 % sure.

So Jeff actually came out this week on September the 15th from a company called TypeSafe AI. And because it never writes, it's fast. Now, why does this matter so much?

Well, a model that writes can make things up. Ask it for an outfit and it might say wear a navy cardigan. But there's no navy cardigan in my closet.

So now the app is stuck. But Jeff, well, Jeff can only pick from the list you give it. So for example, red blazer or leather jacket or even sunglasses, for example, or keep what he has, right?

So every answer is one your app can use right away. And it's literally way, way faster than using a normal AI agent. So how do you reduce your tokens?

Well, when you hand a tiny choice job to Jev, Claude and Jube6 Astra read nothing for that job, zero tokens. Jev still has to read it because somebody has to, but Jev did the same work 99 % cheaper than Jube6 and Astra and ClaudeSonic5. You know, I've tested this on my real jobs.

And let's have a look. So example number one. is my voice mirror, right?

So I say where I'm going and it picks my outfit. I type, I'm giving a keynote tomorrow. That one sentence turns into seven tiny choices, which top, which jacket, which trousers, which shoes, which extras, which room.

And this is even about clothes. Now, Jeff answers all seven together in under a second. Red blazer, white shirt.

navy trousers and a stage behind me. Then I gave those same outfit choices to the big models. 40 requests each.

Jev answered in 413 milliseconds. Quad Haiku took 1486. G6 Astra took 3269.

Quad Sonic 5 took 4545. Now, which is more slow? Of course, it's the models that are thinking.

And for quite a lot of these... steps. You just need something that can decide.

It doesn't need to write anything. It just needs to make a quick decision. And the tiny model can keep up with professors because it's just picking from the choices you've already given it.

Which is how the system works. It's got all these presets and then if we click on something it can basically operate straight away. It can think.

It's the same for example with this system right here. This system allows us to control our browser. Now this is way way faster than using a normal voice agent because the voice agent is going to think this system right here basically just operates in real time.

So for example, we say start mic over here, scroll down the page, open up Google, right? You can easily do it. And we can just stop the mic right there.

Why is it doing that? Because you can see over here that it's got these choices for the different things that we're doing. By the way, this is a free GitHub from...

Moritz Krem, shout out to him for the voice browser system. Really, really cool idea. But this is a whole system.

It can do the same, for example, for emails. It just picks a choice of whether to reply, research, wait, or flag. And you can see here how quickly it can sort through emails as it's running through everything.

So it's just way faster when you're using this. By the way, if you want more training on this sort of stuff, feel free to check out the AI Profit Board. Link in the comments description or go to the AIprofitboard .com.

A great community where you can post questions, get help and support whenever you need to. In the classroom, you can get access to all of my best trainings on this sort of stuff. In the calendar, you can jump a weekly coaching course, ask questions, get help and support in real time.

And this is all available inside the AI Profit Boarding. Link in the comments description or go to the AIprofitboarding .com. Now, the way that I would start implementing this just works in three simple steps, right?

So step number one is you spot the tiny choices. You would look at what your AI agent does all day and find the parts that are really just like a quiz. So if we have a look at, for example, that mirror system, these are the options which answer the question.

And so Jeff can quickly decide through the allowed answers and pick the right ones. For example, with my email system, It can answer, okay, is this email urgent or not?

My website is like, should these two pages be linked to or not? And if the answer could come from a short list, it's a tiny choice and the professor doesn't need to do it, aka the big frontier models don't need to do that. Now, you can actually get your AI agent to figure out this stuff for you.

So, for example, you can give your AI agents like Claude access to the API for Jeff, which you can get from OpenRouter, And then you can get that API, go into Cloud or go into the GPT -6 and say, hey, based on what I'm working right now with, what could we use Jeff for? How can we optimize decision making inside any apps or tools or websites that we've built so that we can save time?

So, for example, you know, you could build a game that just generates itself as it goes based on the decisions you give it. And then the decision making can go to Jeff. It can pick it under a second.

And then your AI agent can keep doing the big work, right? With Claude Astro, whatever you prefer. So planning, writing, fixing things, you know, he can do that.

When he reaches a small choice though, he can use Jeff. And the answer comes straight back and your agent carries on. Now, I like to build with Claude Code.

That's the one I like. But you can do this with any agent. You can tell it use Jeff for this choice and give it your questions.

And then it's got this confidence score as well. So you can set limits. You can say, right, above 60%, you carry on.

Below 60 % on the confidence score. Okay, you get back to me and you tell me and flag it. So that's basically it.

That's the whole system. So you can see from the numbers that I've put out that the cost normally with an LLM is like $0 .2 to $10 per input, right? For JEV, it's $0 .042.

Output is free because there is no real output with JEV. Whereas normally output tokens... with an LLM is five times more than the input cost.

But at the same time, JEV is more for optimizing decisions, which is where you're going to save the most tokens with JEV, right? You wouldn't use this for writing stuff because it can't. And additionally, it's way, way faster.

So not only do you save tokens, but you can also run decisions way faster through this model. Now, it's still in the early stages, but it's pretty exciting. And it's probably one of the biggest releases this year.

when it comes to AI models. And if you want to get more training on this sort of stuff, feel free to check out the AI Profit Boardroom. We have an amazing community and lots of training on Jev AI inside the calendar.

You can drop a week of coach calls, ask questions. In the map, you can meet people in your local area, building out with AI agents like Jev. And also, I answer your questions personally right there too.

Plus, you can direct message me if you need anything. So feel free to get that. Link in the comments description or just go to the AIprofitboardroom .com.

Thanks for watching.

The Hook

The bait, then the rug-pull.

Julian Goldie opens with a blunt reframe: your Claude and GPT-6 Astra usage isn't being eaten by real work, it's being eaten by hundreds of tiny decisions your agent keeps handing to a frontier model that has to re-read the whole conversation just to answer them.

Frameworks

Named ideas worth stealing.

06:20list

The 3-Step Jev Setup

Spot the tiny choices
Write the question and the allowed answers
Send that one step to Jev

A method for finding which decisions inside an AI agent's workflow are closed-list 'quiz' choices, then routing just those to a decision-only model instead of the frontier model handling the whole conversation.

Steal forany Claude Code or GPT agent pipeline with repeated small classification or routing decisions

CTA Breakdown

How they asked for the click.

VERBAL ASK

06:03product

“Link in the comments description or go to the AIprofitboardroom.com.”

Pitched twice (mid-video and again at the close) for the AI Profit Lab community/training. The spoken domain doesn't match the real link in the description (skool.com/ai-profit-lab-7462), which would confuse anyone trying to type it in.

MENTIONED ON CAMERA

06:03productAI Profit Lab ("AI Profit Boardroom") community ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.