All terms
Safety, eval & ops
API (in AI context)
Also known as: model API, inference endpoint
Programmatic access to a model — the HTTP endpoint you POST a prompt to and get a completion back. This is what you actually integrate.
What it means
An AI API is the developer-facing layer of a model provider. You hit an endpoint (api.openai.com, api.anthropic.com, generativelanguage.googleapis.com) with a JSON body containing your prompt, model name, and parameters; you get back a completion. Every chatbot, copilot, and agent shipped in the last three years runs on an API somewhere underneath.
Major providers in 2026: OpenAI API (GPT-5/5.1, o-series reasoning, Sora video), Anthropic API (Claude Opus/Sonnet/Haiku 4.x), Google AI Studio + Vertex (Gemini 3, Imagen, Veo), xAI API (Grok), Mistral La Plateforme, Cohere, plus aggregators like OpenRouter that proxy hundreds of models behind one API. For open-weight models, inference providers (Together, Fireworks, Replicate, Anyscale) expose Llama, DeepSeek, Qwen, etc. behind OpenAI-compatible APIs.
API-level features matter beyond raw model quality. Streaming (token-by-token responses), tool calling, structured outputs (JSON schema enforcement), prompt caching (Anthropic was first; OpenAI and Google followed), batch processing (50% cheaper, async), file uploads, and vision inputs are now table stakes. Picking an API is picking an ecosystem — switching costs are real even with the OpenAI-compatible standard, because tool calling formats, error semantics, and caching behavior all differ subtly.
Example
A simple Anthropic API call: POST to /v1/messages with {"model": "claude-opus-4-7", "messages": [{"role": "user", "content": "Hello"}], "max_tokens": 1024}. You get back a JSON response with the completion plus token usage stats for billing.
Why it matters
When you build with AI, you're building on top of an API contract. Understanding rate limits, pricing, latency characteristics, and feature support of your provider is more important than understanding which transformer architecture they use under the hood. The model is a commodity; the API is the product surface.