Unmasked · Sept 17, 2026 · 23:24 UTC

Union Alpha AI Was Never One Model

It took the internet 33 hours to work that out.

Union Alpha AI was never one model, and that changes how you should pick every AI tool your business uses.

A mystery AI showed up near the top of the coding charts, and the whole AI world spent a day and a half guessing who built it.

Everybody guessing was wrong.

Stick with me, because I'll show you the real scores, the one number they left off the page, and the one line you should check before you send any new AI model your work.

The whole story in one shot: one mask, one name — and underneath it, several minds working as one.

33 hrs
until the mask came off
4-ish
models, one name
262K
tokens of context
5th
stealth launch
Straight from the source

The sources, so you can check it yourself

Official listings, the model card and the reveal ↓
openrouter · live listingPareto by Unbiased → openrouter · the stealth listingstealth/union-alpha → unbiased · model cardPareto 26.9 scores + pricing → unbiased · how it worksHow a Pareto answer gets made → github · their test setupcircuitandchisel/pareto-evals → x · the revealOpenRouter names Union Alpha → the companyCircuit & Chisel →
"Multiple LLMs are engaged in parallel on every request, and their results are synthesized dynamically based on the task."— Unbiased, "How a Pareto answer gets made"
§1Three days ago

A brand new AI model, and nobody would say who made it

It was called Union AlphaBrand newAI modelNear the topfor codingprice: zeroMade by: ???no lab · no logojust a name, and a price of zero
§2A day and a half of guessing

Google? A Chinese lab? A secret OpenAI test?

Thousands piled in that same afternoonGoogle?A Chinese lab?Secret OpenAI test?33 hours latereverybody was wrongit was never ONE modelnot the wrong company — the wrong question
§3Sept 16, 2026 · 14:42 UTC

A new listing appears on OpenRouter

Hundreds of models, one placemodelmodelmodelmodel+ hundredsOpenRoutera shop for AI modelsone loginYour toolplug it inpick a model, plug it into whatever you're using
§4The listing

stealth / union-alpha

stealth/union-alpha
Reads text
and writes it back
Reads images
pictures and thinking together
Uses tools
it does jobs, it doesn't just chat
262,144 tokens in
a few hundred pages of reading at once
131,072 tokens out
in one go
Price: zero
Maker: "Stealth"
specs match the live OpenRouter listing · checked Sept 19
§5The first 37 minutes

Free for a week, and a gold star with no number

Launch day, minute by minute14:42 UTClisting appears+10 minutesOpenCode:"free for a week"+27 minutesCEO posts a chartgold star · no numberso people piled in — and then they started digging
§6The fingerprint test

Every lab's model leaves fingerprints

How you normally unmask a modelOx AlphaAugustFingerprintmatched in 3 daysGLM-5.3-FlashZ.aiUnion AlphaSeptemberFingerprinthow it counts wordshow it words errorsNo matchto anythinglike handwriting — and this handwriting belonged to nobody
§7Sept 17 · 23:24 UTC

OpenRouter posts the name: Pareto, by Unbiased

Union Alphathe maskParetoby Unbiased33 hours after it appeared
the reveal

OpenRouter names it — and says why it ended early

This is the official reveal. Notice the reason: the community's detective work and the demand cut the stealth period short. It was announced as a week. It lasted a day and a half.

§8Why no fingerprint matched

Pareto isn't one AI. It's several, on every request.

What happens when you message ParetoYou sendone messageModel Abig frontierModel Bopen sourceModel CModel D"four-ish"Combineinto oneOne replyone billseveral labs, stitched together, sold under a single name
I asked it: "who made you, and are you one model or several?"● real run · my machine
$ openrouter → unbiased/pareto
"I'm Pareto, an AI service by Circuit & Chisel.
I'm one assistant, but I can't verify whether the service
uses one underlying model or several."
✓ answered in 7.1s · provider: Unbiased

What you're watching: the real answer Pareto gave me this morning. Even the model can't tell you what's underneath it.

§9Stick with me

Four things I'm about to run you through

1 · The real scores
five tests, three of the strongest models
2 · The missing number
the one they left off the page on purpose
3 · The line in the terms
check it on any free AI model before you send it work
4 · The pattern
so you know what to do when the next one appears
§10The thing they were careful to explain

Pareto is not a "router"

Router vs ParetoA ROUTERQuestionGuesseshow hard is it?easy → cheap modelhard → expensive modelPARETOQuestionEvery modelat the same timeOne answer"If a reliable way to do that existed, we'd use it."
"Pareto is not a 'model router.' Routers read the prompt, estimate its complexity, and pick a single model. If a reliable way to do that existed, we'd use it."— Unbiased, "How a Pareto answer gets made"
§11Why this matters for your setup

Your agent sends the same opening chunk, every single step

What an agent really sendsSTEP 1instructionsrulesbackground files+ the new bitSTEP 2instructionsrulesbackground files+ the new bitSTEP 3instructionsrulesbackground files+ the new bitSTEP 4instructionsrulesbackground files+ the new bitCaching = the provider remembers itso you get a discount on that chunkall of it, over and over, on every step
§12The prices they published

Cached input is one tenth of the price

Pareto · price per million tokensInput$2.50Cached input$0.25Output$7.50twenty-five cents instead of two-fifty
live OpenRouter pricing · unbiased/pareto
§13The problem with routers

Swap models halfway through, and the discount is gone

One swap wipes the memory discountStep 1Model Acache warmStep 2Model Adiscount ✓Routerswaps modelStep 3Model Bfull price againback to full price on everything you already sent
§14Why I care about this

My agents share one memory folder

real Agent OS · my machine

What you're watching: the shared memory inside my Agent OS. Claude Code, Hermes and OpenClaw load the same instructions at every step of a long job.

§15Their claim: the discount holds all the way through

Worth testing rather than believing. So I tested it.

cache test · 6,400 tokens of instructions · one growing conversation● real run · my machine
$ same instructions loaded at every step → unbiased/pareto
step 1 · 6,417 tokens in · cached 7 · cost 1.61¢
step 2 · 6,407 tokens in · cached 0 · cost 1.61¢
step 3 · 6,423 tokens in · cached 5,703 · cost 0.34¢ ✓
the discount kicked in by step 3: −79% on the same instructions
three separate one-off requests, same opening: cached 0 every time
What each step cost meStep 11.61¢Step 21.61¢Step 30.34¢inside one conversation it held — one-off requests got no discount
§16Test one · DeepSWE

Coding: a three-way tie at the top

DeepSWE · a coding testPareto74GPT 6 Astra74DeepSeek 4.1 Flash74Fable 5.167Pareto 74 · Astra 74 · DeepSeek 74 · Fable 67
scores published by Unbiased · run by the company that made it
§17Test two · Terminal-Bench

Can it get real work done on a computer?

Terminal-Bench 4.0 · real computer workPareto51Fable 5.156GPT 6 Astra58DeepSeek 4.1 Flash31DeepSeek crashes to 31
§18Test three · MMMU-Pro

Pictures and thinking together

MMMU-Pro · images + reasoningPareto78Fable 5.181GPT 6 Astra87DeepSeek 4.1 Flash77Astra leads with 87
§19Test four · a very hard exam

General knowledge, the hard way

HLE, no tools · very hard general knowledgePareto49GPT 6 Astra54Fable 5.155DeepSeek 4.1 Flash3949 — Pareto's lowest score anywhere
§20Test five · university maths

Maths: 88 — while others fall apart

ArXivMath · university level mathsPareto88GPT 6 Astra91Fable 5.172DeepSeek 4.1 Flash28DeepSeek 28 · Fable drops to 72
§21Hold those in your head

Pareto never wins on its own. It ties once.

Where Pareto finished, test by testtied 1stcoding3rdcomputer3rdpictures3rdhard exam2ndmathseverywhere else: second or third
§22Now look at the bottom of each row

It never falls apart anywhere

Worst day: biggest gap behind the winner, any testDeepSeek 4.1 Flash63 behind · mathsFable 5.119 behind · mathsPareto9 behind · picturesGPT 6 Astra1 behind · examAstra is steadiest — and priced like it ($10 in / $50 out)
§23What several models together gets you

You give up the highest peak. You lift the lowest floor.

Steady vs spiky, on the real scorescodingcomputerpicturesexammathsPareto · steadyDeepSeek · spikysame five tests — one line holds, one line swings
§24Why that matters for agents

One bad step ruins the whole thing

Brilliant at four steps, terrible at the fifth = a messResearchDraftCheckFixPublishsteady beats spiky when the work has stages
§25Credit where it's due

They listed every test they lose on, by name

Terminal-Bench 4.0
Astra 58 · Fable 56 · Pareto 51
MMMU-Pro
Astra 87 · Fable 81 · Pareto 78
HLE, no tools
Fable 55 · Astra 54 · Pareto 49
ArXivMath
Astra 91 · Pareto 88
from Unbiased's own page: "where other models score higher"
§26The number missing from all of it

What does it cost to finish one actual job?

"Measured task costs and a composite score have not been published for this release."— Unbiased, Pareto 26.9 model card
The number that decides cheap or expensivePrice per wordpublishedCost per finished jobnot published?×1×1×1×1four models on every requestwords are the ingredients — the job is the meal
§27The exact problem we work through every week
real Agent OS · my machine

What you're watching: the Agent OS dashboard — every agent and every model in one place, sharing one memory.

Over 3,000 business owners helped · plenty had never used AI before

Get the Agent OS — where trying a new model is a one line change.

A model like Pareto lands every few weeks now. Reading the chart isn't the hard part. Knowing whether to switch the model your business runs on is. The Agent OS plugs Claude, Hermes, OpenClaw and any OpenRouter model into one place with shared memory.

The Agent OS zip file — Claude, Hermes, OpenClaw + any OpenRouter model, one place, shared memory
A 30 day roadmap — for setting it up, plus a video tutorial that walks you through it
Four coaching calls every week — bring your own setup, we'll tell you which model to run for your kind of work
Daily tutorials — the moment a new model lands
Business owners testing models on real client work — just ask someone who tried it yesterday
Beginner friendly — plenty of members had never used AI at all before they joined
Join the AI Profit Boardroom → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description, or search the AI Profit Boardroom on Skool
§28The pattern behind these anonymous launches

OpenRouter has done this five times

The mask → who was underneathQuasar AlphaApr 2025early GPT-4.1Horizon AlphaJul 2025early GPT-5Pony AlphaFeb 2026GLM-5Ox AlphaAug 2026GLM-5.3-FlashUnion AlphaSep 2026Pareto · several modelsevery mask came off — every single time
§29Eleven days, eight, five, six…

Then a day and a half

How long each one stayed anonymousQuasar Alpha11 daysHorizon Alpha8 daysPony Alpha5 daysOx Alpha6 daysUnion Alpha33 hoursand the free window shut the moment the name came out
§30That's the trade

The free week isn't a gift. It's the launch campaign.

Who gets what from a free stealth modelYou geta top tier modelfor nothingThey getthousands of peoplefinding the cracksreal work, real stress test — before they start charging
§31The practical takeaway

The price arrives with the name, without warning

Free listing off · paid listing onAnnouncedfree for a weekIt ran33 hoursPaid goes live$2.50 in · $0.25 cached$7.50 outuse free stealth models — just don't build anything that depends on one
§32The part almost nobody reads

"Doesn't train on your data" is not "doesn't keep your data"

"Prompts and completions may be retained by the provider but are not used for training."— the actual Union Alpha listing
Two sentences. Only one was made.Won't LEARN from your work"not used for training"promisedWon't KEEP your work"may be retained"never promisedthose sound identical — they aren't
§33The plain version

Stealth means logging is switched on

"it's a stealth model, so has logging enabled"— Alex Atallah, OpenRouter CEO, Feb 2026, on an earlier anonymous model
Anonymous window"may be retained"The revealAfter the revealthey say: we keep nothinga good position — it just came after, not during
§34A rule you can use on any free AI tool, forever

Find the retention line

1 · Find the line
not the announcement, not the tweet — the terms on the listing
2 · Train ≠ keep
two separate sentences — check which one you were actually given
3 · Match it to the work
decide by what you're about to send
Rewrite a product descriptionfineA client's private documentsa different decision
§35Who's actually behind this

Two former Stripe engineers. Payments people.

Why the whole pitch is about costStripethe payments companyLouis Amira+ David Noel-RomasCircuit & Chiselthe companyUnbiased → Paretoa smarter way to BUY AITheir pitch: cost per answernot a clever new modelthey didn't train an AI — they built a smarter way to buy it
§36Their own statement after the reveal

More honest than most launches

the statement

"You found us."

This is Unbiased's own post. They say the internet needed one day to pull off the mask. They say some people thought they were testing one new model, that the confusion was their fault and not OpenRouter's, and that Pareto is not a single set of weights.

"The internet needed one day to pull off the mask."
their words
The confusion was their fault
not OpenRouter's
Not a single set of weights
Pareto is a system
More intelligence for less
a new field: many models working as one
Many models working as one.
§37Why it's worth watching

You shouldn't have to ask "which AI should I use?"

"Which AI should I use?"switch monthlyargue in commentsagoniseswitch againOne name. One bill.the picking happensunderneath, every requestwhether they've pulled it off, nobody outside the company can say yet
§38Now the honest pushback

A dot with no number is a prediction, not a result

Every score was run by the company that made itprice →score58563149← no number"expected pricing"if the dot has no number, the number doesn't exist yet
No independent test yetNo public leaderboard entryCharts drawn by the people routing the traffic
§39One more open question

They don't name the models inside the blend

Where your prompt actually goesWhat you sendParetocompany ?company ?company ?company ?fine for most work — for anything private, know it first
§40So how do you find out if it's actually good?

Two models, the same job. Steal the idea.

head to head · same job, same instructions, two models● real run · my machine
$ job: rewrite a product description in under 60 words
unbiased/pareto ........ 4.8s · 61 tokens out · $0.00063
deepseek-v4.1-flash .... 19.5s · 300 tokens out · $0.00017
✓ both usable — Pareto 4× faster, DeepSeek 3.6× cheaper, on THIS job
my job, my scores — a benchmark says nothing about yours
Cost per finished job (my test)Pareto · cost0.063¢DeepSeek · cost0.017¢measure the job, not the million words
Their testing setup is public ↓
github · MIT licencecircuitandchisel/pareto-evals →
§41The simple way to think about all of this

Three things to know

One · cost per finished job
not per million words — words are the ingredients, the job is the meal
Two · how steady it is
agents break at their weakest step, not their best one
Three · what happens to what you send
kept or not · trained on or not · named company or anonymous
§42Where this shows up day to day

Same weekly report? The cached price decides your bill.

Same long instructions, every timeHeadline price1.61¢ a runCached price0.34¢ a runwhy your costs don't match the pricing page
numbers from my own cache test above
§43The speed of all this

From eleven days to under two

Mystery model → everyone knowsOx Alpha6 daysa technical fingerprintUnion Alpha33 hourspeople comparing notesThe next onelessstill true next month, when the next one shows up
§44If you want to use it right now

The anonymous version is off. This one is live.

unbiased/pareto  ·  or "pareto" on their own platform
live listing · openrouter.ai/api/v1/models● live API · Sept 19
$ curl openrouter.ai/api/v1/models | grep pareto
id ............ unbiased/pareto
context ....... 262,144 tokens · max output 131,072
input ......... text + image · tools ✓
price ......... $2.50 in · $0.25 cached · $7.50 out (per million)
✓ listed under Unbiased — stealth/union-alpha is switched off
Still in betaEvery signup approved by handFree onboarding hour with a purchase
§45What I'd actually take from all of this

Own the system. Rent the model.

OLD WAY
a weekend

Reading every announcement

  • A new model lands, you start from scratch
  • Re-paste your instructions into a new tool
  • Rebuild your workflows around it
  • The free window shuts before you finish
  • Next month, do it all again
NEW WAY
10 minutes

A setup built for swapping

  • A new model lands, you change one line
  • Your agents keep all their instructions
  • Same memory, same dashboard
  • Test it on your real work the same day
  • Keep what wins, drop what doesn't
real Agent OS · recorded during the Ox Alpha window

What you're watching: the model dropdown in my Agent OS. Last month it was Ox Alpha. Switching is one dropdown, and everything around it stays the same.

YOUR SYSTEM · you own thisMemoryInstructionsAgentsModelrented · swap memodels keep arriving — the system is the part you keep
Own the system. Rent the model.
§46If that sounds like it needs a technical background

It doesn't. Three beliefs to drop.

Wrong: I need a technical background to build a setup like this.
Right: Most of the people I work on this with had never touched an AI tool before they started. The zip file and the 30 day roadmap do the heavy lifting.
Wrong: I need to pick the single best AI model and stick with it.
Right: Models arrive faster than anyone can test them. The winners built a setup where trying a new one takes ten minutes instead of a weekend.
Wrong: If a model tops the chart, it's the right one for my business.
Right: A benchmark tells you how a model does at the benchmark's job. Run your real work through two models, and keep your own scores.
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 them in their own words.

Read the 158-page wins doc →
§47Your move
real Agent OS · my machine

What you're watching: the agents inside my Agent OS. Point them at a different model and they keep every instruction.

The setup where switching to something like Pareto is a two minute change

Get the Agent OS — your agents keep their instructions, whichever model you point them at.

Claude, Hermes, OpenClaw and any OpenRouter model plugged into one place with shared memory. That's exactly what the Agent OS inside the AI Profit Boardroom gives you.

The Agent OS zip file — plus the 30 day roadmap to get it running and a full video walkthrough
Daily updates — as we improve it
Four coaching calls every week — bring your own model tests, we go through your setup live
Daily tutorials — how to use new models to get more customers, instead of just playing with them
A prompt library — built for agent workflows
A member map — find people near you doing the same thing — there's always someone online
Join the AI Profit Boardroom → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description, or search the AI Profit Boardroom on Skool
THINKING IT? "Doesn't running an Agent OS burn a fortune in tokens?"

No. Everyday work runs on free local models and free APIs, and the frontier work drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and the Agent OS plugs straight into it.

Inside the Boardroom there are full token-efficiency tutorials too, so you learn to cut usage right down.

§48One last thought

Several AI models, pretending to be one

One maskUnion Alpha33 hoursthe next one: lessthe next one will take less
Union Alpha was several AI models pretending to be one. The next one will take less.
Keep exploring · more guides
GLM 5.3 + Hermes Agent: same brain, way better training
An agent that orders your stock, two desks, seven models and a prompt file worth deleting
Hermes /loop: the agent that watches for you
DeepSeek Harness: everything is a plugin
Claude AI SEO: Rank #1 Fast
The Full Claude AI SEO Course
Hermes Bot Mode: the free alternative to Grok Bot
Two seats, three projects, one skill, a live agent map and a VPS to rebuild
Grok 4.6 AI SEO
← Browse all guides  ·  By topic: Hermes · Claude · AI SEO · Free Claude Code · Build your own