Buy Intelligence.
Sell Inference.

A model-blind exchange for AI output.

We standardize hundreds of AI models into three benchmarked tiers and certify sellers against public standards — so you buy trusted intelligence without managing models or vendors. Anyone with compute can sell: AI clusters, gaming GPUs, even idle edge devices.

Inference is becoming a commodity. Performance is the standard we grade it by — and futures will make its cost predictable.

Private demo · play credits (real mechanics & real AI answers, no real billing yet). Developers & AI agents: see intellex.rasonx.com/llms.txt.
What's For Sale

Intelligence, in three grades

Tier 1· PhD-level▲ 1.8%
~$5 / megatoken (1M tokens)
Min speed ≥ 15 tok/s

For your hardest problems — deep reasoning, math, multi-step analysis. Frontier-grade answers, without paying for a frontier brand name.

Graded against:
Claude Opus 4.xGPT-5.xGemini 3 ProDeepSeek R1
Tier 2· College-level▼ 0.9%
~$0.50 / megatoken (1M tokens)
Min speed ≥ 30 tok/s

For everyday work — answering questions, summarizing, drafting, light coding. Reliable help at roughly a tenth of Tier 1's price.

Graded against:
Claude Sonnet 4.xGPT-5 miniGemini 3 FlashLlama 3.3 70BDeepSeek V3
Tier 3· High-school-level▲ 4.2%
~$0.02 / megatoken (1M tokens)
Min speed ≥ 60 tok/s

For simple, high-volume tasks — formatting, sorting, tagging, pulling facts out of text. So cheap it's effectively free.

Graded against:
Llama 3.x 8BMistral SmallQwen2.5 7BGemma 2 9B
These name-brand models are the ruler, not the seller — trading stays blind, you never see which model fills your order.
Example prices — real prices are set by sellers competing with each other, not by us. Every grade is enforced by certification: how the grading engine works →
How the tiers are set · the Model 500

The tiers aren't arbitrary. Every model we grade is ranked into one index — the Model 500 — and the tiers are rank bands of it: Tier 1 is the top of the field, Tier 2 the next band, Tier 3 the rest. Each band's floor is that tier's bar, so as stronger models are ingested the ranking reconstitutes and the bars rise — a tier always tracks the frontier, like the S&P 500 tracks the market. See the live index →

For teams already spending on inference

Stop paying for the brand name

Tier 1
The same frontier-grade reasoning you'd buy from a top provider — certified to the bar, billed at the cleared market price.
vs. a frontier model's list price
Tier 2
Everyday workhorse capability — same bar, filled by whoever can deliver it cheapest.
vs. a mid-tier model's list price
Tier 3
High-volume bulk work — so cheap it changes what's worth automating.
vs. a small model's list price
If you already route through OpenRouter or pay the big providers directly, you're paying a brand-name premium for a level of capability. Buy the tier instead — certified to the same bar (quality and speed), priced by a live market, fully blind.
Why a tier beats a model

Buy the bar once — it keeps up for you

Never migrate models again

Wire up Tier 1 once and you ride the frontier. Every new state-of-the-art model — the next GPT, Claude, Gemini, or DeepSeek — is graded into the tier as it ships, and the basket that defines Tier 1 moves up with it. Your one API call keeps getting smarter on its own. No swapping model names, no re-testing, no migration every time the frontier moves — the standard tracks it for you.

Always on — automatic failover

No single model is load-bearing. If the seller filling your tier slows down, errors, or drops offline, the exchange instantly fails over to the next certified seller at that same tier — you still get your answer, at the grade you paid for. It's the way a power grid keeps the lights on when one plant trips: you buy the tier, and the market keeps it delivered.

How It Works

How a trade happens

1

Ask for what you need

Pick a tier — Tier 1, Tier 2, or Tier 3 — and send your question. (For programmers: it's one API call, the same format as OpenAI's, with "grade" where the model name used to go.) Or post a bigger job on the job board with a fixed reward.

2

The market finds your seller

Your question goes to whoever sells that level of intelligence at the lowest price right now. Posted jobs work like an auction in reverse: sellers bid lower and lower to win the work, and the cheapest one gets it.

3

You get the answer, they get paid

Payment happens instantly, and every fraction of a cent is tracked. On jobs, you pay only the winning bid — the rest of your money comes straight back. Who did the work stays secret: you're paying for the quality of the answer, not for a name.

Who It's For

Built for both sides of the trade

$For buyers

  • Pay for the answer, not the name. The same work often costs far less when no brand name is attached.
  • Easy to plug in. If your software already talks to OpenAI, one small change makes it talk to IntellEx instead.
  • Start in seconds. Sign up, get a free starter balance, and see exactly where every fraction of a cent goes.
  • Set a budget, keep the change. Post a job with a fixed reward — sellers compete to do it for less, and you keep whatever's left over.

For sellers

  • You bring the model. Selling means running your own AI model that's already up on hardware you control — locally with Ollama, LM Studio, or vLLM on your own GPU, or any hosted OpenAI-compatible endpoint. The exchange sends each prompt to your endpoint and relays the answer; it doesn't host or download a model for you.
  • You must pass the grade. Your endpoint is tested against the tier's public bar (quality and speed) before it can sell, and re-checked while it's live — so buyers can trust the tier.
  • You set your price. Raise it, lower it, pause, or walk away any time. The seller app handles the busywork for you.
  • Nobody knows it's you. Buyers see a code name and a price — never your name, your model, or your setup. Stored keys are encrypted and used only at the moment of a sale.
Why hide who's selling?

When buyers can see brand names, big names get to charge extra just for being big names. When they can't, sellers can only win on the two things that actually matter — price and how good the work is. That's the whole idea. And "how good" isn't taken on faith: the grading engine tests every seller with real exams, and takes the badge away if they slip.

Enter the exchange →
You'll be asked for the access password.