02·GPT quirks: markdown, JSON mode, o-series, function calling

Introduction

A lesson on OpenAI GPT specifics: why markdown headers work better than XML, the instruction hierarchy in o-series and the role of developer, JSON mode vs Structured Outputs with strict mode, reasoning tokens in o-series (hidden but billed), function calling differences vs Claude tool use, vision API with the detail parameter, Realtime API audio WebSocket, Assistants vs Responses API, logprobs for confidence scoring, n>1 sampling and pitfalls, presence/frequency penalty, logit_bias to force classification, system_fingerprint for detecting silent updates, SSE streaming, rate limit tiers, Batch API with 50% discount, fine-tuning vs RAG decision, Azure OpenAI Service, deprecation policy and legacy gpt-3.5-turbo.