Understand tokens, inference, and sampling. Call the API with factual, creative, and instruction prompts. Sweep temperature and compare outputs. Learn top_p and top_k — know when to adjust each and when to leave them alone.
Count tokens before sending. Every provider charges per token — know your model’s pricing, context window limit, and how to estimate cost per call. Use a tokenizer library (tiktoken for OpenAI, Anthropic’s token counter) to verify what you’re actually sending.
Also learn: how to handle 429 rate-limit errors (exponential backoff), when a smaller/cheaper model is good enough, and how batching API calls (OpenAI Batch API) cuts cost by 50% when latency doesn’t matter. You’ll hit all three on day one.
Know the landscape: OpenAI (GPT-4o, o-series), Anthropic (Claude), Google (Gemini), Meta (Llama), Mistral, Cohere, and DeepSeek are the providers that matter today. Each has different strengths — don’t marry one.
!!! tip “Start here” Pick one provider, make one API call, print the response. That’s Module 1 done at the minimum. Everything else in this module is depth.
OpenAI docs · Anthropic docs · Google Gemini docs · The Illustrated GPT-2 · Tiktokenizer