The AI credits glossary.
Credits, tokens, quotas, and the ideas between them. Each term includes a practical example and a useful next step.
API
An application programming interface: a defined way for software to request a service or exchange data. An AI application uses an API to submit inputs and receive model outputs.
In practice: A document tool sends a request to a model service rather than asking a person to copy text into a chat window.
Explore the related guide ↗API credits
An allowance or accounting unit used to pay for eligible API usage. It may be a prepaid monetary balance, a promotional award, or a platform-defined unit; the issuer defines what it covers.
In practice: A balance can fund requests without representing a fixed number of tokens across every model.
Explore the related guide ↗API key
A credential used to authorize access to an API. It is not a credit balance, and possessing one does not establish that billing, model access, or quota requirements are satisfied.
In practice: Keep production keys in a controlled server environment, not in the public code of a website.
Explore the related guide ↗Automatic reload
A billing feature that purchases additional credits when defined conditions are met. Its trigger and purchase controls should be reviewed separately from the application’s usage policy.
In practice: A team can permit a reload while still limiting the number of expensive tasks its own application starts.
Explore the related guide ↗Batch processing
Submitting a group of tasks for asynchronous processing instead of requiring an immediate response to each one. Eligibility, completion windows, and pricing depend on the provider and endpoint.
In practice: Overnight catalog enrichment can be evaluated separately from an interactive support conversation.
Explore the related guide ↗Cache hit
A request component that matches reusable cached material under the service’s rules. It may receive different billing treatment from uncached input, but that behavior must be verified for the chosen service.
In practice: Measure reported hits instead of assuming that two prompts with similar meaning share a cache entry.
Explore the related guide ↗Cache miss
Input or work that cannot use the relevant cache for a particular request. A realistic forecast should include misses even when a controlled test achieves substantial reuse.
In practice: Run a stress scenario with less cache reuse to see whether a feature remains affordable.
Explore the related guide ↗Context window
The amount of content a model can consider within the applicable request limits. The precise accounting of input, output, and other components depends on the model and interface.
In practice: A long document and a conversation history can both contribute to the context the application sends.
Explore the related guide ↗Cost per accepted task
The chosen cost boundary divided by the number of results that meet a defined acceptance standard. This is an evaluation metric, not a provider’s universal billing unit.
In practice: If a test costs $5 and yields 1,000 usable results, the measured component is $0.005 per accepted task.
Explore the related guide ↗Credit expiration
The point after which an allowance can no longer be used under its terms. Purchased balances and promotional awards can have different validity rules, so record them separately.
In practice: Review unused funds before a deadline rather than assuming every balance remains available indefinitely.
Explore the related guide ↗Embedding
A numerical representation of content used for tasks such as similarity search and retrieval. Generating, storing, and querying embeddings can add separate components to an AI application budget.
In practice: A retrieval system can find relevant passages before passing selected context to a generative model.
Explore the related guide ↗Fine-tuning
Additional training that adapts a model using a selected dataset. Its availability, training costs, and subsequent serving costs are separate questions from the price of a standard inference request.
In practice: Budget for dataset preparation and evaluation before assuming fine-tuning is a cheaper replacement for prompting.
Explore the related guide ↗Free tier
An access level that permits certain usage without a direct charge within defined conditions and limits. It is not necessarily a monetary credit or a guarantee of production capacity.
In practice: A provider may allow bounded evaluation of eligible models while requiring paid access for other work.
Explore the related guide ↗Inference
Running a trained model to produce an output from an input. The serving arrangement determines how that work is metered and who is responsible for the infrastructure.
In practice: A hosted inference request and a model running on your own machine have different cost boundaries.
Explore the related guide ↗Input tokens
The token units representing content supplied to a model under its metering rules. User text can be only part of the input; instructions, history, and tool results may also matter.
In practice: A short question attached to a long document can involve much more input than the question alone.
Explore the related guide ↗Large language model (LLM)
A model trained to process and generate language, with capabilities that vary by model and configuration. Choosing an LLM for a product requires task-specific evaluation rather than a universal ranking.
In practice: Evaluate a support assistant on representative customer questions and its own acceptance requirements.
Explore the related guide ↗Open-weight model
A model whose trained parameters are made available under specified terms. Access to weights does not remove license obligations or the cost of hosting and operating inference.
In practice: Gemma hosting requires a serving environment even when model weights are available to download.
Explore the related guide ↗Output tokens
The token units generated by a model under the service’s accounting rules. Visible answer text may not describe every billed output component for every model or operating mode.
In practice: Inspect reported usage when comparing a short classification result with a longer reasoning workflow.
Explore the related guide ↗Pay-as-you-go
A usage-based charging arrangement rather than a fixed entitlement to unlimited service. When payment occurs, which balances offset it, and which caps apply depend on the provider’s plan.
In practice: Usage may be charged through a funded balance or an invoiced arrangement, depending on the service.
Explore the related guide ↗Prepaid balance
Funds committed before eligible usage and drawn down as the service applies charges. A prepaid balance should not automatically be treated as an instantaneous hard spending cutoff.
In practice: Keep reload permissions and application-side task limits distinct from the visible balance.
Explore the related guide ↗Prompt caching
Provider-supported reuse of eligible prompt context. It differs from an application returning a previously saved answer, and its matching, lifetime, and billing rules are service-specific.
In practice: Measure initial and repeated requests separately when evaluating the benefit of stable document context.
Explore the related guide ↗Rate limit
A constraint on throughput over a defined interval, such as requests or tokens per minute. It is different from a remaining credit balance or a monthly spending limit.
In practice: A funded application may still need to queue a burst of requests instead of sending them all at once.
Explore the related guide ↗Startup AI credits
A promotional award intended to offset eligible development or operating resources under a program’s conditions. Approval, coverage, duration, and customer contribution are issuer-specific.
In practice: Maintain an unsubsidized cost forecast even while an approved award pays part of the actual bill.
Explore the related guide ↗Token
A unit used by a model’s tokenizer to represent content. A token is not universally equivalent to a word, character, API request, financial asset, or fixed amount of credit.
In practice: Measure token usage for the actual language and format rather than relying on one universal word ratio.
Explore the related guide ↗Tokenized API credits
An ambiguous phrase that may describe platform metering units or a claimed token representation of service access. The label alone does not establish transferability, authorization, or redemption rights.
In practice: Verify the issuer and actual service terms before treating a token-like balance as spendable API access.
Explore the related guide ↗Definitions have a scope.
These are editorial explanations, not a replacement for an issuer’s contract. Billing units, cache behavior, model access, and grant conditions remain provider-specific. Examples are illustrative.
Read the official source notes ↗