OpenAI + Webtwizz

GPT-4o-mini is the workhorse. At $0.15/M input tokens, it's the right default for classification, extraction, routing, simple chat, RAG synthesis on small context. If you don't have a strong reason to use something else, use this.

GPT-4o is the upgrade for quality-sensitive work, sales-facing chatbots, code suggestions in your app, anything where users will notice mistakes. ~$2.50/M input. Still fast (sub-second first token), still cheap enough for high-volume.

o1 / o1-mini are reasoning models, slow (10–60s per response), expensive ($15+/M), but solve multi-step problems mini and 4o get wrong. Use for math, complex code generation, strategic planning. Don't use for chat, users won't wait.

GPT-5 (when available in your tier) is the new top of the lineup. Webtwizz defaults the OPENAI_MODEL env var to gpt-5.4 if you don't set it. Override for cheaper/faster work.

Routing pattern: in any chatbot or AI feature, classify the input first with 4o-mini ("is this question simple or complex?"), then route to 4o or o1 accordingly. Webtwizz can scaffold the routing layer if you describe the rule.

Three OpenAI features that materially change what your app can do, none auto-scaffolded on integration connect, but all easy to ask the AI for:

Streaming responses. Without streaming, your chatbot user waits 5–15 seconds staring at a loading spinner. With streaming, tokens appear word-by-word. Implementation: pass stream: true to the chat completions call, return a ReadableStream from your Next.js route, consume it with EventSource or fetch in the client. Ask "make this chat route stream the response" and Webtwizz scaffolds it.

Function calling (tools). Let GPT trigger your app's code. "user wants to update their email" → GPT calls your updateEmail() function with the right arg. Implementation: define a tools array in the chat completion call, handle tool_calls in the response, execute the function, return the result. This is how you build agents that DO things, not just chat.

Structured output (JSON mode). Force GPT to return valid JSON matching a schema you provide. Use response_format with a Zod schema. Critical when GPT's output feeds another piece of code: no more "GPT wrote prose around my JSON" parsing nightmares.

All three are pre-installed in the openai npm package. Webtwizz generates the patterns when you ask; it doesn't pre-scaffold them because most apps need them shaped differently.