Pricing and fees

Each reply is paid one of two ways. A key’s method is chosen when the key is made; the chat apps switch from the balance in the header.

• Account credits. A new account starts with a $5 signup credit, spent before any credits you buy and removed 90 days after it is granted. Buy more by card under Billing in the portal, any amount from $5 to $500; purchased credits do not expire. With a saved card, automatic replenishment buys more ($20 by default) when your settled balance falls below a threshold you set ($5 by default), up to a monthly cap ($100 by default); credit held for a reply in progress does not trigger it. Billing shows the range of credit prices per million tokens that a Katara Cortex reply is charged at, since each reply costs the rate of the model that answered, and the price of a picture and a web search. Credits pay for Live keys and the chat apps, on Katara’s own models and on Katara Community while it is on; Test model calls use the environment’s test credits instead.
• USDC, from your wallet through a spending approval or a prepaid deposit. Requests paid in USDC or with credits can also run on independent providers in Katara Community, unless you turn it off; a reply paid with credits costs the same credits either way. With it on, Community providers are tried first, and a reply from one may take longer to start.

Paying with credits

A request paid with credits holds enough of your available credits to cover its maximum cost while it runs. When it completes, the exact cost of what was delivered is charged, never more than the hold, and the rest is released; a failed request is charged nothing. When the available credits cannot cover the maximum, the request is refused with 402 insufficient_credits.

A web search costs 0.006 credits, charged only when it succeeds and the request ends with an answer, and an image has one price per picture.

The rest of this page covers USDC.

Paying with USDC

Prices are per million tokens, separately for input and output, in USDC. Each provider sets its own for each bundle it serves. A reply from one of Katara’s own models costs the price Katara lists for that model, the same number as its credit price. GET /v1/models shows Cortex’s price ceiling. The Deep and Standard models change price on set dates; Models lists both prices for each.

No Katara Community provider can ask more than 2.40 USDC per million input tokens or 12 USDC per million output tokens, on any model and however the request is paid. A provider may also set a minimum price per job, at most 0.022119 USDC, the cost of 4,096 input and 1,024 output tokens at those rates. A request a provider serves never costs more than those rates applied to the larger of itself and that reference request. Katara’s own models charge their listed prices, which can be higher: the Deep model’s are from 1 December 2026.

What you pay

A reply costs input tokens × input price + output tokens × output price at the price of the provider that served it, or that provider’s minimum job price when it is larger. The number is on the reply and in your usage history. Two percent of every settled reply goes to Katara; it is included in the price you see. A reply paid per request from a wallet settles in its own transaction, and a fixed part of its charge also goes to Katara to pay for that transaction: 0.012 USDC, or 0.018 USDC with the upto scheme, which takes two transactions. Katara’s share never exceeds the charge, so a reply charged less than that goes to Katara whole; you still pay only the charge.

How it is charged

Replies are paid from a prepaid balance held in the settlement contract, not straight from your wallet. When that balance cannot cover a request’s maximum cost (the input estimate plus max_tokens at the offer price), Katara first moves a small chunk, one dollar at launch, from your wallet into it through the approval you gave. If your wallet cannot fund even the shortfall, the request is refused and nothing is charged.

When the reply is complete and the provider’s signed receipt checks out, the exact cost is debited from that balance. Debits are grouped into one transaction every ten seconds, which is what keeps the fee at 2% instead of gas eating it.

If a provider fails mid-reply, nothing is charged for that attempt.

Paying per request from a wallet

Programs that pay with a wallet signature on each request, without an account, settle each reply in its own transaction. Those requests carry a minimum charge so that a very short reply does not cost more to collect than it earns. It is 0.025 USDC for katara/cortex and for embeddings and images, 0.05 USDC for katara/cortex-genius and 0.01 USDC for katara/cortex-smart. Only this kind of request can name the last two (Choosing a model). Replies above the minimum are charged their settled cost as usual, and the reply’s cost_usdc is the charge, the minimum included. Replies paid from an account balance have no minimum. Pay per request with x402 shows how to quote, sign and send such a request.

Spendable balance

Spendable is your prepaid balance plus what the approval can still move in from your wallet. The wallet panel shows both parts. When the budget follows your balance, spendable equals your balance.

Taking money out

The prepaid balance is yours. Ask for a withdrawal in the wallet panel and it pays out after a one-hour delay, minus any replies that were still being settled when you asked. While a withdrawal is pending, new requests are refused so nothing can be served against money that is leaving.

Test

In Test the token is test USDC on Avalanche Fuji, which has no value, and the faucet gives 5 a day per account. Charging, the fee and withdrawals work as they do in Live. Providers on Avalanche Fuji set their own prices, so a Test price is not a Live one. Paying per request without a key works in Test too, with the same minimum charges, paid in test USDC.

An agent’s Test Compute on a credit budget is the exception: it is charged in Account Credits, at its real cost. On a wallet or x402 budget it is paid in test USDC. See Networks.