Tested Sep 20, 2026 · four real jobs · Claude Haiku 4.5, Claude Sonnet 5 + GPT-6 Astra

Jev AI: How to Reduce Claude + GPT-6 Astra Tokens by 99%

Here's how to reduce your Claude and GPT-6 Astra tokens by 99 percent with Jev AI.

Right now your big AI models are doing hundreds of tiny jobs they don't need to do, and every one eats your tokens.

Jev is a tiny AI that takes those jobs, answers in under half a second, and tells you how sure it is.

I raced it against Claude and GPT-6 Astra on 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 to set it up, and two bonus moves for your everyday chats.

Stick with me, because one of those bonus moves made my agent forget its own rules, and the fix was one number.

Stop paying a professor to sort your mailTiny choices"which shoes?""is this urgent?"Jev picksunder halfa secondOnly when surelow number =ask a personThe big AIreads zero forthose jobsFour real races on my own machine
The actual sources ↓
§1What a token is

Tokens are the words your AI reads

Tokens in plain wordsA tokena small piece of a wordYour planhas a token limitMore readingslower answersHit the limityour agent stopsLess reading means faster agents that run longer
§2The hidden problem

It reads the whole chat again, every time

Every new message re-reads everything before itMessage 1readsone pageMessage 20reads thewhole chapterMessage 50reads thewhole bookYour message is tiny. The reading is huge.
§3Real examples

Three tiny messages. Huge reading.

What each message made the big model read"Can you add this one too?"124,124 tokens"Do you have the guide on it as well?"196,239 tokens"how's it going?"222,190 tokensFrom my own real agent chats
§4The real waste

You're paying a professor to sort your mail

Two very different kinds of workTHE PROFESSORbig thinkingPlanning a projectWriting a guideFixing something trickyBuilding an appALSO GETStiny choices"Is this email urgent?""Which shoes?""Should these pages link?""Is this a good sales message?"The tiny choices don't need a professor
That's like paying a professor to sort your mail.
§5What Jev is

Jev is an AI that only picks

How you talk to JevWhat's going onyour situationA questionin plain wordsA list of answersyou write themJev picks one+ how sure it isIt never writes. It picks.
§6Why a list matters

It can only pick from your list

Why the answer list mattersWRITING MODELcan guess"Wear a navy cardigan"You don't own oneNow your app is stuckAnd it took secondsJEVpicks from your listRed blazerLeather jacketCamel overcoatOr: keep what he hasEvery answer is one your app can use
§7What the 99% means

For these jobs, the big AI reads nothing

Where the 99% in the title comes from0tokens Claude and Astra read, once Jev takes the job99.5%cheaper than GPT-6 Astra on the same work4real jobs I tested it onMeasured on four real jobs, on my own machine
§8Race one · the mirror

One sentence becomes seven small choices

the real mirror app on my machine · real Jev calls

What you're watching: I ask for a keynote outfit and every slot fills at once. Then "lose the jacket" changes one thing and leaves the rest.

§9Race one · the times

The same 40 choices on four models

real run on my machine · replayed at reading pace

What you're watching: the same 40 outfit choices run on Jev, Claude Haiku 4.5, GPT-6 Astra and Claude Sonnet 5 — the middle time for each, then the score.

§10Race one · the tokens

Claude read 74,923 tokens. Then zero.

40 outfit requests, two waysBIG MODEL4.6 secondsClaude read 74,923 tokensGPT-6 Astra read 45,897For 40 outfit requestsAll of it counts on your limitsJEV0.4 secondsClaude reads 0GPT-6 Astra reads 0Same 39 out of 40 rightYour limits are not touchedThe small choices moved off the big model
§11Race one · bonus trick

It even marks its own work

the real mirror app on my machine · real Jev calls

What you're watching: I ask for a podcast outfit. Jev's first pick scores 1.9 out of 3, so the mirror asks again and the new look scores 2.7.

§12 Want this inside your own agents?

Get the AI agent system I run, with Jev built in.

The helper at the door plugs into the same place your agents already live.

The Agent OS — one place to plug in your Claude, your Hermes and your OpenClaw, and add your own workflows
The zip file, a 30-day roadmap and a video walkthrough, with daily updates
Coaching calls where we go deep on cutting tokens with Jev, and you can ask about your own setup live
Daily tutorials on making Claude and GPT-6 Astra agents run longer on the same limits
3,900+ business owners, some who had never used AI before, and members already building with Jev
A member map so you can find people near you, and someone is always online
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description · used in 38 countries
§13Race two · the inbox

60 emails, sorted in 2.5 seconds

real Jev run on my machine · replayed at reading pace

What you're watching: 60 test emails land in four piles — reply, research, wait, or flag to me — each with how sure Jev was.

The tiny model sorted the most emails correctly.
§14Race two · the scoreboard

The tiny model got the most right

real run on my machine · replayed at reading pace

What you're watching: the same 60 emails on four models — the middle time for each, then how many each one sorted correctly.

§15Race three · sales messages

Spot the message sent to the wrong business

real Jev run on my machine · replayed at reading pace

What you're watching: 40 sales messages get scored. Eight were written for the wrong kind of business — Jev flags every one.

§16Race three · the scoreboard

Everyone got 40 out of 40

real run on my machine · replayed at reading pace

What you're watching: the same 40 sales messages on four models — times, then right answers.

§17Race four · my website

1,542 yes-or-no calls in under six seconds

real Jev run on my machine · replayed at reading pace

What you're watching: Jev reads my website page by page and decides which pages should link to which — and leaves some pages alone.

§18Race four · the scoreboard

Jev was 20 times faster than Sonnet

real run on my machine · replayed at reading pace

What you're watching: a 40-page sample of the same job on four models — the middle time for each page.

For those four jobs, the big models read zero.
§19The final scoreboard

All four races: 99.5% cheaper than Astra

real run on my machine · replayed at reading pace

What you're watching: all four jobs added together — how much cheaper Jev was than each big model, and how many tokens the big models read once Jev takes over.

§20How to do it · step 1

Step 1: spot the small choices

Small choices hiding in my own apps"Which outfit?"my mirror app"Which button next?"my voice browser"Is this email urgent?"my inbox agent"Which page links here?"my website agentAnything with a short list of answers
§21How to do it · step 2

Step 2: write the question and the answers

a real question from the mirror app
QUESTION
Which shoes should he wear for this request?

ALLOWED ANSWERS
keep            the request doesn't call for new shoes
white sneakers  casual, smart-casual, filming
brown loafers   smart-casual, dinners, chinos
chelsea boots   night out, cold weather
black oxfords   formal suits, stage, weddings
neon runners    gym, running, workouts
slides          beach, pool, home
§22How to do it · step 3

Step 3: send that one step to Jev

Only the small choice goes to JevYour agenthits a smallchoiceJevpicks in underhalf a secondYour agentcarries on withthe answerThe big model still does the big work
§23How to do it · the safety number

It knows when to do nothing

the real mirror app on my machine · real Jev calls

What you're watching: "did I lock the front door" and "how many members joined this week" — the mirror stays still. Then "swap the shoes for the chelsea boots" — it changes one thing.

§24How to use the number

High number: go. Low number: ask me.

How sure Jev was"did I lock the front door?"100% sure · not about clothes44 of Jev's 54 right answersabove 60% sure60 emails · all 6 missesevery one below 60% suremy line60%Every email it got wrong sat under my line
Every email it got wrong sat under my line.
§25Bonus move one

Send easy messages to a smaller model

Two kinds of messageQUICK JOBsmall model"Can you add this one too?""Do you have the guide as well?""I've topped up now.""how's it going?"HEAVY JOBbig model"Create guides on two tools""Set up https and handle it""What can I build with Agent OS?""Review your skill, update the guide"Real messages I typed to my agents
§26Bonus one · the test

Jev sorted 98 of my real messages

real run on my machine · replayed at reading pace

What you're watching: 98 real messages I typed to my agents, sorted by Jev into quick job, not sure, or heavy job. The counter adds up the tokens the quick jobs would have made the big model read. Longer or private messages are hidden.

§27Bonus one · the check

Claude agreed on 37 of 40

Checking Jev's sorting40quick jobs Jev was 80%+ sure about37 of 40Claude Sonnet 5 agreed58stayed with the big modelA second opinion from the big model
§28Bonus one · the saving

Quick jobs were 36.2% of the reading

Where the title's 36% comes from36.2%of everything the big model read4.3Mtokens, spent on quick jobsMore than a third of the reading, on messages like "how's it going?"
§29Bonus one · the safety line

Only act when it's 80% sure

How sure Jev was"how's it going?"100% sure · quick"Can you add this one too?"97% sure · quick"my to-do list disappeared, get it back"82% sure · quick ✕my line80%One of these was a miss
§30Bonus move two

Throw away the old news

Real calls from my mirror-app chatTHROWN AWAY · "re-run the tests"99% sure · they passed hours agoTHROWN AWAY · "deploy the guide"99% sure · it's doneTHROWN AWAY · "test eight phrases"98% sure · old resultKEPT · "read the design helpers"the agent was still using themJev only saw what each result was for, and its first few lines
§31Bonus two · watch it tidy

174 results judged in 5.2 seconds

real run on my machine · replayed at reading pace

What you're watching: every old tool result from the chat that built the mirror app gets a call from Jev — keep, not sure, or throw away — with its size and how sure Jev was. Then the before and after token counts from Claude's own report.

§32Bonus two · the safety line

Only throw away when it's 50% sure

How sure Jev was about throwing things awaythe script I'd written14% sure · KEEPmy saved notes23% sure · KEEPmy line50%a test that passed99% sure · throw awayBelow the line, it stays
§33Bonus two · my mistake

I cut 51% and my agent forgot its rules

real run on my machine · replayed at reading pace

What you're watching: on another long chat, my first try threw away everything Jev marked as old news. The four biggest things it dropped were my agent's own rule book — and Jev was under 50% sure about every one.

§34Bonus two · the memory test

30 questions: 30, then 24, then 27

Could Claude still answer 30 questions about the job?Nothing thrown away30 / 30Threw away everything · 51% fewer tokens24 / 30Only when 50%+ sure · 15% fewer tokens27 / 30The sure-ness line is what keeps your agent smart
§35Be clear about this

The tokens move. They don't vanish.

What the 99% does and doesn't meanSTAYS THE SAMEthe readingJev read 24,904 tokens of emailThe big models read about 6,000 to 9,000Jev's format is wordierThe reading doesn't disappearCHANGESwho reads itClaude and Astra read zeroYour big-model limits are saved99% cheaper than AstraAnswers in under half a secondOnly bonus move two makes the pile itself smaller
If someone shows you a huge token cut, ask what the agent forgot.
§36Where to start

Start with the tiny choices

The order I'd do them in1st · small choiceseasiest, nothing to lose2nd · easy messagesbiggest saving: 36%3rd · old newsneeds the 50% safety ruleEasy first, careful last
§37Is this too hard?

Each helper is one question

Everything you have to writeJob 1"Which shoes?"Job 2"Quick job or heavy job?"Job 3"Keep, or throw away?"Plus one numberwhen is Jev sure enough to act?Plain words. No code from you.
Wrong: “I'd need to be a developer to cut tokens like this.”
Right: Each helper is one plain question, a short list of answers, and one number for when to act. Your coding agent does the wiring.
§38About limits

Hitting limits isn't just how it is

Where my own tokens were going222,190tokens for "how's it going?"36.2%of reading was quick jobs21%of my mirror chat was old news I could dropMost people never look
Wrong: “Running out of tokens is just what it's like to use AI agents.”
Right: A third of my reading was quick messages, and a fifth of one chat was old news. Both are fixable.
§39The honest limits

What I haven't proven yet

Six honest limitsLink qualityI timed it, I didn't grade itTest emailswritten for the test, not my real inboxOne run eachbig models can vary run to runBonus movesmeasured on saved chats, not live3 of 30 facts losteven with the safe ruleJev is in betafive days oldKnow these before you copy me
§40Should you wait?

What you learn stays with you

Five days from release to real numbersSep 15Jev comes outSep 20944 calls onmy own chatsWhat staysknowing where yourtokens goThe helper may change. The lesson doesn't.
Wrong: “I'll wait until the tools settle down.”
Right: A faster helper will come along. Knowing which choices are small and what your agent can forget carries straight over.
Don't take my word for it

Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries.

Read the 158-page wins doc →
§41 Your move

Make your agents run longer on the same limits.

Three small questions in front of the big model. Inside the AI Profit Boardroom you get the system they plug into, and help setting it up.

The Agent OS zip — plug in Claude, Hermes and OpenClaw, then add workflows like the Jev helper
A 30-day roadmap + video walkthrough, with daily updates
Four coaching calls a week — bring your own chats and we'll look at where your tokens go
Daily tutorials on cutting tokens for Claude and GPT-6 Astra agents so you save time
3,900+ business owners and a member map to find builders near you
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description · used in 38 countries · new tools added the week they ship