AI APIs and How to Use Them

AI APIs and How to Use Them

If you want to add AI capabilities to an application, you almost certainly start with an API. AI APIs let you call a language model over HTTP, send a prompt, and get a response — no ML knowledge required. Understanding how they work makes you a far more effective builder.

The Basic Pattern

Every major LLM API follows a similar pattern:

1. Authenticate with an API key
2. Send an HTTP POST request with your messages
3. Receive a JSON response with the model's output

The request body typically looks like a list of messages with roles (system, user, assistant). The response includes the generated text, token counts, and metadata.

Most providers offer official SDKs for Python, TypeScript/JavaScript, and other languages that wrap these HTTP calls. In practice you use the SDK rather than constructing raw HTTP requests.

Pricing is almost always per token: you pay for the tokens you send (input) plus the tokens the model generates (output). Output is typically more expensive.

The Main APIs

OpenAI: the most widely used. Ships the GPT model family, including reasoning models with adjustable reasoning effort. Large ecosystem of tooling built around it. Good default choice for most use cases.

Anthropic (Claude): strong long-context handling, careful output quality. Often preferred for document analysis, careful reasoning, and coding tasks.

Google (Gemini API): via Google AI Studio (developer) or Vertex AI (enterprise). Deep Google ecosystem integration; context window sizes are now comparable to the other frontier labs rather than a standout advantage (see callout).

Together AI, Fireworks, Groq: API access to open-weight models with fast inference. Often cheaper than frontier model APIs for tasks where open-weight models are sufficient.

Ollama (local): runs models on your own machine. No API cost, no data leaving your system — but you provide the compute.

As of October 2026: context windows of roughly 1 million tokens are now standard across the frontier labs — it's no longer a Gemini-specific advantage.

Practical Tips

Set a system prompt: shape the model's behaviour before the user's message. Don't skip this.

Handle errors gracefully: APIs have rate limits, occasional outages, and timeout errors. Build retry logic with exponential backoff.

Watch your token usage: log input and output token counts. Costs can scale unexpectedly when prompts or contexts grow.

Stream responses: most APIs support streaming (receiving the response token by token as it's generated). This makes UIs feel much faster — users see text appearing rather than waiting for the full response.

Use structured outputs: modern APIs support JSON mode or function calling to get structured data back reliably. Far better than asking the model to "return JSON" in free text and trying to parse it.

Cache aggressively: if you send the same system prompt or context in many requests, most providers now offer prompt caching that reduces cost and latency significantly — including Anthropic and Google's explicit prompt caching, and OpenAI's automatic prompt caching that applies without any extra configuration.