One ledger for language, voice, image, and video APIs. Compare billing units and source-linked rates, then model the work your product will actually run.
API video generation, with clip-based and per-second rates kept distinct.
View pricing →Speech generation APIs compared by character, token, or audio-duration billing.
View pricing →Text models compared using separate input and output token prices.
View pricing →Image APIs with resolution, quality, and token-based billing made explicit.
View pricing →Artificial intelligence is a broad field: systems can classify information, make predictions, find patterns, or generate new material. A model is only one part of an AI product. Data sources, instructions, software tools, safety checks, and the way people use its output all affect the result.
This guide introduces the main ideas behind common AI APIs, how to test an AI feature, and why the cheapest model is not always the best fit. It is written for curious readers and developers who want a clear starting point rather than a model leaderboard.
Most API calls are inference, not a fresh training run. The application decides what context to send and what checks happen before a response reaches a person.
Machine learning is one approach within AI. Instead of encoding every decision as a fixed rule, developers train a model to learn patterns from examples. A classifier might label a support message; a recommendation system might rank items; a generative model might draft text or create an image. Each solves a different kind of problem and needs a different success measure.
Google's introduction to machine learningTraining fits a model from data; inference is using an already-trained model to respond to a request. An API request can include instructions, conversation history, documents, or media. Language models process text in tokens: a token can be a word, part of a word, punctuation, or another piece. Token counts therefore differ from word counts and affect usage limits and many API bills.
OpenAI API concepts: prompts, models, and tokensLanguage models can draft, summarize, classify, or extract structured data. Image, speech, and video systems handle different inputs and outputs; multimodal models work across more than one kind of media. Check the actual endpoint for supported formats, file limits, context windows, and output controls. Compare equivalent tasks with the same quality, resolution, and duration.
Retrieval-augmented generation (RAG) is a pattern in which an application searches its own document collection, selects relevant passages, and supplies them to a model with the question. Semantic search can find related meaning even when wording differs. The model can then draft an answer grounded in those passages, but retrieval can miss the right source or surface outdated material.
OpenAI guide to retrieval and semantic searchA prompt gives instructions and context for one request. Retrieval supplies relevant external material that can be refreshed separately. Fine-tuning, when a provider supports it, adapts a model using examples for a repeated task or output style. Tuning is not a live search engine: changing facts still need a current source. Start with a clear prompt and representative tests before taking on a training workflow.
Google Cloud overview of model tuningWith tool or function calling, a model can return a structured request such as looking up an order or searching a database. The application validates the request, checks permissions, runs the function, and sends the result back to the model. The model does not automatically have access to every system. Keep credentials and authorization in your application, and require human approval for consequential actions.
OpenAI guide to function callingA benchmark or polished demo cannot tell you whether a model fits your product. Build realistic examples, include difficult edge cases, and define what a good result means before testing. Compare models with the same inputs and settings. Track correctness, consistency, response time, and cost together. For creative tasks, use a clear human rating rubric because there may be several acceptable answers.
Google's guide to measuring ML project successGenerative systems can produce incorrect, incomplete, or biased output, and model behavior can change between versions. Verify important claims against reliable sources. Before sending private or regulated information to an API, read the provider's data-retention, training, and access terms. For consequential uses, plan human review, permissions, fallback behavior, and a way to report problems.
NIST's Generative AI Risk Management ProfileFor a text API with separate input and output rates: input tokens / 1,000,000 x input rate, plus output tokens / 1,000,000 x output rate. A monthly budget should also account for average request size, output length, retries, caching, tools, and traffic peaks. Image, voice, and video services may use different units such as image, character, second, or clip, so a token-only estimate will not cover them.
Retrieval does not guarantee that sources are complete or correct. Show the source passages, test search quality, respect document permissions, and make it easy for a person to verify an answer.
Keep side effects explicit. A tool that reads public data has a different risk from one that changes an account, sends a message, or moves money.
Use a repeatable check instead of choosing from a demo alone.
Google: What is machine learning?OpenAI: Counting tokensOpenAI: RetrievalOpenAI: Function callingGoogle: Measuring successNIST: Generative AI risk profile
API Ledger is a pricing comparison, not a model quality ranking. Listings link to provider documentation and show when their displayed model, billing unit, and rate were reviewed. Read how pricing data is collected or explore the API pricing guide.