Vex Inference

You choose a model in VEX Desktop, and VEX runs it for you. You pay in $VEX, from a balance you deposit in the app.

1. The hold. Before a request runs, VEX holds the most it could cost. A request can never spend more than your balance.
2. The charge. When the model answers, you are charged what the model provider billed, plus the 15% VEX share, converted to VEX at a 60-minute average price (TWAP). If the provider reports no bill, because you stopped the request, the connection dropped or the provider failed mid-answer, you are charged an estimate instead: the input tokens VEX counted and every output token received, at the model's listed prices, plus the VEX share, never more than the hold. A request the provider rejects before it answers costs nothing.
3. The release. The rest of the hold returns to your balance at once.

The models and their prices are live at inference.projectvex.ai/v1/models (opens in a new tab). Prices are in USD per 1 million tokens. You pay in VEX, converted at the 60-minute TWAP.

Prices include the 15% VEX share.

Here a token is a piece of text a model reads or writes, not the $VEX token. Input is what the model reads, output what it writes. Cache read is input the provider serves from its cache at a lower price, and cache write is input stored in that cache for the next request.

Prices include a 15% VEX share, and all of it goes back into VEX.

For every $1 of model cost:

Burns run in batches; each one is a public transaction. The burned VEX leaves the supply.

The treasury's address: 0xe341f3da256C38356bce4Afd456d7fa36E356E94 (opens in a new tab)

Deposits

• You deposit VEX from the app.
• Up to $25 of a deposit is usable within seconds, as provisional credit.
• The deposit is final after about 16 minutes on Robinhood Chain.

The deposit address, on Robinhood Chain (chain 4663): 0xf36aed1c11A3140594609969DC6646c9F8a981C5 (opens in a new tab)

Withdrawals

• A withdrawal waits through a 12-hour cooling.
• You can cancel it while it cools.
• It goes only to a wallet linked to your account.

Prompt caching passes through to you. When a large prompt repeats, the provider's cache prices apply, and the charge follows them.

Claude models need an explicit cache breakpoint, and VEX Desktop sends one. OpenAI models cache automatically.

Measured on 2026-10-22, from Poland to the gateway in Azure Central US. The gateway's own time is what VEX adds to a request before the model starts; the model's time and the network path are not in it.

Method: one campaign of paired requests per model, timed from each response's Server-Timing header. Of the 1,403 checked charges, 1,378 were compared with the cost the provider billed. The other 25 were requests that failed without a bill from the provider; each was charged an estimate and checked as one.