Glossary
This page defines the terms that have a meaning of their own on Liquid Inference. It does not define words that every LLM API uses, such as model, token and provider. Each entry links to the page with the full rule.
API key
A key looks like sk-liquid-<key-id>-<secret>. Each key has one scope:
spend: send requests, and do everything below.read: read receipts, balance and the market.list: read market data only.
The console shows the secret one time only. See API keys.
Auction
The steps that pick a provider for your request. The exchange removes each offer that fails one of your limits. Then it ranks the rest and gives the job to the best offer. See Purchasing rules.
Auto
A router model, liquid.auto, that chooses the model for you. It compares a published score for each model with the estimated charge. Add a preference to the name: premium, balanced or economy. Auto needs a price bound, such as a price cap. See Auto Router.
Axis
A word after a colon in the model name, such as deepseek:cheapest. It sets how the exchange ranks the offers. The axes are smartest, fastest, cheapest, premium, balanced and economy. An axis changes only the order. Each offer must still pass your limits. See Axis routing.
Balance
The money that a workspace can spend. It counts paid funds and credit. A workspace under an organization spends the balance of the organization.
Basket
A list of models that you name in one request. Send architect.basket as the model, and list the models in exchange.candidates.models.allow. The offers of all the listed models compete. See Purchasing rules.
Book
All the offers that providers have posted now. It is also called the order book. See Market data.
BYOK (bring your own key)
A way to send requests through your own account at a cloud, such as Anthropic or Amazon Bedrock. The cloud bills you for the tokens. The exchange charges only its fee. A key that you connect serves only your workspace. See Bring your own key.
Charge
What you pay for a job. It is the price of the tokens that the job used, plus the exchange fee. It is never more than the price cap. If the job uses more than its cap, the provider pays the difference. See Receipts.
Cost
The number that the exchange uses to rank offers by default. It is the estimated charge plus the wait cost. It is not the amount that you pay. See Price and wait.
Cost tag
A label that you add to a request with the x-liquid-tag header, such as a team or project name. Reports can group your spend by it. The exchange never sends it to a provider. See Purchasing rules.
Creator
The company that made a model, such as OpenAI or DeepSeek. Many providers can run the model of one creator, so the creator is not always the provider.
Deadline
A time limit that a provider promises in its offer. There are two: the time to the first token, and the time to the end of the answer. If the provider misses a deadline, the receipt shows missed. You still pay for the job. See Market data.
Deployment
One model as one provider runs it, with stated terms such as precision, region and limits. Its symbol is the provider name and the model name, such as acme/gpt-oss-120b. Use the symbol as the model name to buy from that deployment only.
Estimated charge
The price that the exchange expects for your request, before the job runs. The exchange uses it to rank the offers. You pay the real charge, not the estimate.
Fallback
A move of a job to the next offer. If the winning provider fails before it writes any output, the exchange can try the next offer. After the first output, the exchange never changes the provider. This rule prevents an answer from two providers.
Fee
The amount that the exchange adds to the charge of each job. For a job priced in tokens, it is the larger of 5% of the provider's charge or a small floor for each token. A job that fails before any output pays no fee. See Fees.
Guardrail
A set of rules for what is never allowed: for example, which providers or models can serve a call, how much a key can spend, and which sensitive data a call cannot send. Each workspace has a default guardrail that covers every key. You can add guardrails that cover all keys or only the keys that you pick. A key follows every guardrail that covers it. A request cannot widen a guardrail. Only owners and admins can change one. See Guardrails.
Hold
Money that the exchange keeps aside from your balance while a job runs. Before the first token, the exchange holds the price cap. After the job, it charges the real cost and returns the rest to your balance.
Job
One request and the work to answer it. Each job has one job ID, in the x-liquid-job-id response header, and one receipt.
Market
All the offers for one model, from every provider. A market is liquid when it has three or more independent suppliers. See Market data.
Nano-USD
One billionth of a US dollar. The API gives exact amounts of money as whole numbers of nano-USD, in fields that end in _nanos.
Offer
A provider's standing price for one model, with the deadlines that it promises. An offer stays on the book until it expires or the provider changes or removes it. A job pays the rates of the offer that won it, as they were at that time.
Organization
An account that pays for the workspaces under it. Their jobs are charged to the balance of the organization. Use one when a company has several teams and wants one balance. See Organizations.
Output limit
The maximum number of tokens that an answer can have. You set it in the request, for example with max_tokens. The price cap assumes that the answer reaches this limit.
Preset
A saved set of models, a ranking strategy and limits, with a name. Make one on the Presets page. Then send its name as the model. A value in the request replaces the saved value. See Purchasing rules.
Price cap
The most that a job can cost, fee included. The exchange fixes it before the first token. It is the worst case: the answer reaches its output limit, and no cache saving counts. The x-liquid-cap-usd response header shows it. To set your own cap, send exchange.cap_usd.
Prompt caching
A provider stores the start of a prompt that repeats, such as a long system prompt. A cache write stores it. A cache read uses it again at a lower rate than normal input. A multi-turn chat or an agent loop gets the saving with no setup. See Prompt caching.
Provider claim
A statement that a provider makes about how it handles your data or which audits it passed, such as SOC 2. The exchange shows each claim as a badge. You can require a claim, so that only providers that hold it serve your calls. See Provider claims.
Reasoning tokens
Hidden steps that some models write before the answer. You pay for them as output. On a reasoning model, the first reasoning token counts as the first token for the time to first token.
Receipt
The record of one job. It shows which offer won, the tokens used, the charge, the fee and the speed. A receipt does not change after settlement. See Receipts.
Refusal
An error that the exchange returns when no offer meets your limits. The error is 503 no_eligible_offer, and the response gives the reason that each offer failed. A refused request costs nothing. See Purchasing rules.
Reservation
A monthly pool of tokens or units on one model from one provider, at a fixed monthly fee. You ask providers for quotes and accept one. You use the pool through a preset. Each reservation requires zero data retention. See Reservations.
Response caching
The exchange stores a whole answer. An identical request later gets the stored answer back, and no provider runs. A stored answer costs nothing. It is off until you send x-liquid-cache: true. See Response caching.
Retention
Whether a provider keeps the text of your prompt. Send exchange.privacy.retention as none to use only offers that promise not to keep it. The provider makes this promise, and the exchange does not check it. To use only promises that the exchange checked, require the zero_retention claim. See Data handling.
Role
What a person can do in a workspace. There are three roles:
- Owner: everything that an admin can do, and can manage other owners.
- Admin: manages API keys, presets, guardrails and members.
- Member: reads them, but cannot change them.
See Workspace members.
Router model
A model name that covers many models, such as liquid.auto. The offers of all its models compete, and the exchange ranks them. GET /v1/exchange/routers lists each router model and its rules.
Score
A number that a third party, such as Artificial Analysis, publishes for a model. Auto and the smartest axis use it. A score is about the model. It does not measure the provider that answers your request.
Settlement
The last step of a job. The exchange charges the tokens used, pays the provider, keeps its fee and returns the rest of the hold to your balance. After settlement, the receipt is final.
Spending limit
A maximum amount that one API key can spend in a day, week, month or its full life. When the key reaches the limit, its requests fail with 402 budget_exhausted, even if the balance is higher. See API keys.
Strategy
The rule that ranks the offers that pass your limits:
price: the lowest cost wins. This is the default.fastest: the highest output speed that the exchange measured wins.smartest: the highest published score wins.auto: the best score for the estimated charge wins.
See Purchasing rules.
Supplier
An independent source of offers. Linked accounts, and resellers of the same source, count as one supplier. The exchange counts suppliers to tell if a market is liquid.
Wait cost
The money value of the time that you wait for an answer. It is the predicted wait multiplied by the wait value. The default wait value is 200 nano-USD for each millisecond, which is $0.72 for each hour. The exchange adds the wait cost to the estimated charge to rank offers. You do not pay it. See Price and wait.
Web search
Add :web to the model name to search the web before the model answers. The search has a fixed fee and the cost of the search service. This feature is in beta and can change without notice. See Axis routing.
Workload
A test that finds the cheapest model that does your task correctly. You give a task and some cases with the answers that you expect. A run sends each case to up to four lanes and ranks them by correct answers and cost. See Workloads.
Workspace
The account that holds your balance, your API keys and your receipts. Sign-up makes one for you. A workspace can be on its own or under an organization. See Workspace members.