Today's concept: model vs API vs app - the layers of the AI stack
Right now I'm using an app (Claude Code) on a flat subscription, talking to the exact same model that, called straight through the API, bills $2 per million input tokens and $10 per million output. Same brain, two different price tags, because you're paying for different layers of the stack.
Model = the brain. API = the phone line to the brain. App = the whole business built on that phone line.
1) The model is the engine. The neural network with billions of trained weights that does the thinking (Claude Sonnet 5, say). Like a car engine it does the work, but it has no steering wheel or dashboard: it doesn't remember you, can't read your repo, can't run anything. Training one takes data, GPUs and months.
2) The API is the phone line. An endpoint you send text to and get text back from, billed per token (roughly 4 characters each). This is where the model becomes rentable: no GPUs, no weights on your disk, just a key and a bill. It's also where the knobs are: model ID, output length, tools, caching.
3) The app is everything built on that phone line. Claude Code, Cursor, your company's support bot. Memory, retrieval over your docs, tool calls, permissions, the loop that lets the model act more than once: all app-layer code wrapped around the API.
The layers also move at different speeds, and that's what bites. Model IDs are versions: on the Claude API every ID is a pinned snapshot, so claude-sonnet-5 serves the same weights for as long as it exists. That's how you decide when you upgrade, instead of being upgraded on someone else's schedule, because even a smarter model can quietly shift formatting, tool-calling or tone and break a tuned prompt.
Notice where the interesting problems sit. Context engineering (what goes in the context window, what stays out). Prompt caching, the app-side use of KV cache reuse: a cache hit bills at 10% of the normal input rate, the biggest cost lever most teams never pull. MCP, the open standard letting one tool integration work across Claude, ChatGPT, VS Code and Cursor. All app layer.
Which is the job: as an agentic AI engineer you'll almost never touch the model layer. You wire API calls, tools, memory, evals and permissions around a brain someone else trained.
Quick check before you scroll: Anthropic ships a brand-new, smarter Claude model tomorrow. Does every app built on the Claude API instantly get smarter too?
Full breakdown + the answer: frankduah.me/learnings/2026-08-08-model-vs-api-vs-app-the-layers-of-the-ai-stack
New here? I post a bite-size AI / ML concept like this every day - follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#AI #LLM #AIAgents #MachineLearning
The answer
No. Every API call names a specific model ID string. Nothing changes until a developer edits their code to point at the new model - apps don't auto-upgrade just because a better model exists.