An AI chatbot is the product you talk to — a chat window with memory, guardrails, and a personality — while a large language model (LLM) is the underlying prediction engine that generates the words; the chatbot is the car, the LLM is the engine.
Chatbots like ChatGPT and Claude are built on top of LLMs, which is why two apps can feel completely different even when they run on similar technology. Understanding the split matters because it tells you where to aim your complaints: when an answer is wrong, that's usually the model; when the app forgets your last message or refuses a request, that's usually the chatbot layer wrapped around it.
The mechanism is worth slowing down for, because it explains almost every odd behaviour you'll see. An LLM does one job: given everything so far, predict the next chunk of text. It has no database of verified facts and no built-in truth check.
A chatbot adds several layers on top — a system prompt that sets tone and rules, a memory store that replays earlier turns, and often a retrieval step that pastes relevant documents into the prompt before the model answers. That last layer is the big one. It's why a support bot can quote your exact refund policy: someone wired the policy text into the request, not because the model memorised it.
Context length sets the ceiling on how much of that material fits at once. According to our AI tool database, OpenAI's flagship assistant is GPT-5.5 with a 1M context window, and Anthropic's Claude Opus 4.8 also carries a 1M context window. In practice, a million tokens means you can feed in something like a long contract, a codebase slice, or hundreds of pages of chat history and still have room for the model to reason over all of it — which is what makes multi-turn memory feel continuous rather than goldfish-like.
Here's a decision rule you can actually apply. Count your distinct intents — the number of genuinely different things a user might ask. If that number is small and your answers must be word-for-word compliant (think: "what are your opening hours?", "how do I reset my password?"), build a scripted bot.
It's cheaper, auditable, and cannot improvise a policy it shouldn't. If your intents run past roughly fifty, or users phrase things unpredictably, or you need summarising and rewriting rather than lookup, use an LLM with retrieval grounding. Concrete example: a hospital appointment line has maybe twenty intents and legally fixed wording — scripted wins.
A software company's developer-questions inbox gets thousands of phrasings of "why is my build failing?" — an LLM grounded in the docs wins, because no script author can anticipate "it worked yesterday and now it's angry." The tipping point isn't the model's intelligence; it's whether your answer set is closed or open.
Where this gets messy: the split isn't always clean. Many products are hybrids — a scripted menu for billing, an LLM for everything else — and users can't tell which they're talking to, which is exactly why they get frustrated when the smart bot suddenly can't handle a simple transfer.
Costs also diverge sharply, and this is where the reference material earns its keep. According to our AI tool database, DeepSeek's API runs at $0.435 per 1M input tokens and $0.87 per 1M output tokens, a pay-as-you-go rate described as roughly 12x cheaper than GPT-5.5 — while ChatGPT Plus is $20/mo and Pro is $200/mo, and Claude Pro is $17/mo on annual billing or $20/mo monthly, with Max starting from $100/mo.
So the honest answer to "which is cheaper" depends on volume: a low-traffic internal tool may cost pennies on an API and waste money on a subscription, while a heavy daily user gets more value from a flat plan. The catch is that cheap tokens don't buy you accuracy — a grounded retrieval setup costs engineering time, and a scripted bot costs maintenance every time a policy changes. Neither is free; they just bill you in different currencies.