How $Tokens pays for the models.
Hold 10,000 $Tokens. Chat on this site. Wire the same catalog into your own product. Creator fees on the coin buy the inference. This page is the whole mechanism, in our words.
The short version
Tokens is a simple trade. People trade $Tokens on pump.fun. Those trades throw off creator fees. We collect the fees into a treasury and spend them on model replies. If your wallet holds at least 10,000 $Tokens, you can use the catalog. You do not stake. You do not lock. You do not burn a personal credit bar after every message.
The wallet is the account. The balance is the pass. The treasury is the bill payer. Public pages show whether fees coming in can cover replies going out.
That is the product. Not a subscription with extra steps. Not a promise that utility might appear later. The cost already exists. Someone has to pay DeepSeek, and every other upstream that sits behind a row in the catalog. $Tokens is how that bill gets funded.
Inference is a bill
Ordinary software mostly ran on servers you already rented. One more click did not open a new invoice. Models are different. Each reply spends GPU time. Agents that sit open all day spend more. A product that chats on behalf of users spends more still.
So every AI product ends up answering the same question: who pays when the model talks? Usually the answer is a credit card on an API dashboard, or a monthly seat. Tokens tries a different answer. Market activity on the coin creates the cash. Holders consume the compute that cash bought.
A card on an API dashboard, or a monthly seat. Each reply is your bill.
Creator fees fill a treasury. Holders spend the compute that treasury bought.
Crypto is useful here for one boring reason. A token can both move in a market and sit in a wallet we can check on-chain. Fees can land in a treasury we publish. Usage can be counted. Runway can be a number instead of a slide.
How the money moves
Every swap of $Tokens on pump.fun accrues creator fees in SOL. A keeper looks at that wallet about every 2 minutes, claims what it can, and treats the value as prepaid inference credit. One hundred percent of claimed fees are AI credits. Nothing is skimmed off to add liquidity.
Anyone can swap $Tokens. Any size.
Claimed value becomes AI credit. All of it.
One Solana wallet. Address is public.
10,000+ $Tokens. Tokens stay in your wallet.
Chat on the site, or call the API from your app.
One upstream invoice. The pool pays it.
No personal credit bar. Rate limits protect the shared pool.
- 1The coin trades
Swaps on pump.fun create creator fees.
- 2The treasury claims
A bot sweeps on a 2-minute clock.
- 3Credits land
100% of claimed fees become AI credit.
- 4Holders spend it
Chat here, or call the same models from your app.
There is no mystery remaining after that. Upstream providers invoice in dollars. Each request has a measurable cost. If the coin is quiet, the pool thins. If holders hammer the catalog, burn rises. If fees outrun burn, runway grows. The loop is the product. It either works in public or it does not.
The door is a number
Hold 10,000 $Tokens. That is the whole access rule. Below it, the models stay closed. At or above it, you get the same catalog as every other wallet that clears the line. Holding a million does not buy a bigger slice than holding 10,000. We are not splitting a fixed pie pro rata. We are checking a threshold for a shared service.
You can read the site. The models will not spend the pool for that wallet.
A bigger bag does not buy a bigger slice. Sell under the line and it stops.
Tokens stay in your wallet. Sell under the line and access stops on the next balance check, about 60 seconds. Buy back over the line and it returns just as fast. We count whole tokens only. Dust below one full token does not count. The connected Phantom wallet is the one we measure, not a leftover session from a wallet you already disconnected.
Until the contract address is set on this deployment, the gate reads as not launched. Drop the mint, redeploy, and holders plus the 10,000 check turn on from that one value.
How you use it
On the site: connect a wallet, sign once to prove you own it, then click any model. Signing does not move funds and does not grant spend permission. It only binds a session cookie to that address.
No API key. The session cookie is the account. Signing never moves funds.
The key is hashed. It still dies if the wallet drops under the hold line.
From your own product: mint an optional API key on the keys page and call POST /v1/chat/completions the way you would call OpenAI. Change the base URL. Keep the rest. The key is hashed with SHA-256 before we store it. A lost key is replaced, not recovered. The key does not override the chain. If the wallet behind it drops below 10,000 $Tokens, the next request is rejected.
You do not need a key to chat here. Keys exist so you can drop this catalog into bots, backends, and internal tools without stuffing a wallet signature into a cron job.
Every catalog row is a live terminal for that model. The list is text models. Replies are fulfilled through DeepSeek so the prepaid pool has one upstream bill instead of a pile of vendor accounts.
Watch the pool
A mechanism like this should not hide in a private dashboard. Prepaid credit, burn over the last 24 hours, runway, requests, claims, and the treasury address sit on Transparency. The same snapshot is GET /api/state. History is GET /api/series.
Treasury, claims, requests, burn.
The numbers the site itself reads.
Runway, prepaid, the 10,000/day cap in public.
If a figure cannot be proven, we show zero rather than a story.
Runway is pool value divided by recent spend. If nobody is chatting, runway is not a boast, it is unused credit. If spend spikes, runway shrinks in public. Coverage is the same idea over a week of burn. None of those figures are a guarantee. They are arithmetic on what already happened.
Market cap, when the mint is live, is pulled from the open market and refreshed on the homepage about once a second. Holders are counted on-chain. If a number cannot be proven, we show zero rather than a story.
Ceilings
Shared credit is not infinite credit. Limits exist so one wallet cannot empty the pool in an afternoon, and so the whole site cannot light the treasury on fire in a day.
| Per wallet, per day | 10,000 requests |
| Per wallet, per minute | 120 requests |
| High-price models | 8 per minute when output is at or above $5 per million tokens |
| Per wallet spend | $2 of inference cost per day |
| Whole site | $30 of inference cost per day across every caller |
The daily request cap is the one you will feel first. The dollar caps exist because request count is a bad proxy for cost. A short ping and a long reasoning pass are not the same bill.
What a bot does, and what a person still does
Be precise about automation. The keeper can watch the creator wallet, claim fees, and write the public ledger. The site can check balances, hash keys, count requests, and proxy chat. Burn, prepaid, and runway can be computed without anyone touching a spreadsheet.
Creator fees arrive as crypto. DeepSeek bills in dollars. Moving value from the Solana treasury into the prepaid inference account is still an operator step. There is no on-chain covenant that cryptographically forces every lamport of fees into model spend. The wallet is a normal treasury wallet. You can see it. You can watch the credit. You should not pretend a smart contract is wiring the GPU invoice by itself.
Claims creator fees and writes the public ledger.
SOL and USDC sit on-chain where anyone can look.
The prepaid account that actually answers chats.
Fees in and usage out are measured. The top-up in the middle is still a person.
That gap is not a secret. If markets are going to fund software, the seams should be visible. Better payment rails may close more of this later. Until then, the honest version is: fees in are measurable, usage out is measurable, the top-up in the middle is still a person.
The catalog can change
Tokens is not a bet on one model name. The row you click this month may be gone next year. Prices fall. New labs show up. A reasoning model that looks unbeatable gets replaced. That is fine. The financial loop does not need to predict the winner.
We publish a live catalog and an OpenAI-shaped chat interface. Today that catalog is text. Every reply is fulfilled through DeepSeek so there is one prepaid bill to watch. If the menu of models changes, the rule stays: $Tokens generates the capital, the treasury buys the compute, holders use it.
What this is not
It is not free AI. GPUs still invoice. We moved the payer. You experience a door instead of a shrinking personal meter. The treasury still gets the receipt.
It is not a lockup, a vault, or a points program. If you sell, you leave. If you buy back, you return. No multiplier, no boost for holding longer, no hidden tier above 10,000.
It is not a promise that volume continues. If $Tokens stops trading, fees stop. If holders consume faster than fees arrive, runway falls. Those are not edge cases around the design. They are the design, running in public.
The bet
For years, token utility meant a maze of extra verbs: stake, boost, vote, farm. The useful version might be smaller. The coin opens something people actually want. The coin's own trading pays to keep that thing on.
With $Tokens, that thing is inference. Hold 10,000 $Tokens and the catalog opens. Fees fill the treasury. The treasury pays DeepSeek. You chat, or you ship a product on the same API. The ledger shows whether the fees can keep up.
If it works, a market can fund the intelligence its holders use. If it does not, the numbers will say so without a blog post to translate them. That is the experiment Tokens is running.