What Is an LLM? Large Language Models Explained — and Why Thai Costs More

Short answer: an LLM (Large Language Model) is a language model trained on an enormous amount of text until it can predict, one piece at a time, what token should come next — repeating that until a full answer forms. ChatGPT, Claude, and Gemini are not LLMs themselves; they are applications wrapped around one. First, though: in Thai search results "LLM" is two completely unrelated things.

What an LLM actually is

"Large" means two things at once: the training data is large (vast amounts of web text, books, and code) and the model itself is large (an enormous number of parameters — the numbers the model adjusts as it learns). What an LLM does is one thing only: predict the next token. It does not look an answer up in a database. That single mechanism is why it can draft an email, summarise a contract, translate, and write code — and it is exactly why it can be confidently wrong, because "the next token that sounds right" and "the next token that is right" are not the same thing.

• •Not a search engine — it consults no source while answering unless you wire tools to it (web search, RAG, MCP)
• •It did not memorise what it read — what remains inside is statistical pattern, not a copy of the documents
• •It knows nothing after its cutoff — your prices, your stock, your refund policy: it cannot know them unless you feed them in

How an LLM works: characters to tokens to an answer

The model reads neither characters nor words. It reads tokens — text fragments fixed when the model was trained. Google's own documentation states it plainly: "For Gemini models, a token is equivalent to about 4 characters. 100 tokens is equal to about 60-80 English words." Roughly four characters per token, in English. Hold onto that number — Thai is about to be measured against it. One cycle looks like this:

1. 1.Split everything into tokens — system prompt, chat history, and the latest question, all as one block
2. 2.Score the probability of the next token against everything currently in context
3. 3.Pick one token, append it, and loop back to step 2
4. 4.Stop at the end of the answer or at max tokens — the text trickling onto your screen is that loop, not an animation

Why Thai costs more — measured, not guessed

API billing counts tokens, not characters, and almost every tokenizer was trained predominantly on English. Add the fact that Thai is written without spaces between words, and the model has a much harder time finding word boundaries. Thai text ends up chopped into smaller pieces and billed as more tokens than English carrying the same meaning. I measured this with OpenAI's own tiktoken library, using five real Thai shop messages paired with faithful English translations, across two tokenizers: o200k_base (the current GPT-4o/GPT-5 family) and cl100k_base (the GPT-4 / GPT-3.5 era).

• •Repeated against 93,606 characters of Thai from this blog: 2.68 characters per token — close enough to the small sample that the figure isn't cherry-picked
• •ปัญญาประดิษฐ์ (13 characters) = 8 tokens, while artificial intelligence (23 characters) = 3 tokens
• •ใบเสร็จรับเงิน = 7 tokens, while receipt = 1 token
• •But ประเทศไทย = 1 token — newer tokenizers have learned common Thai words. The good news is the gap narrowed from 3.37x to 1.46x in a single tokenizer generation

In baht: what a LINE OA chatbot actually costs per month

Say a shop runs a chatbot on LINE OA handling 5,000 messages a month. Each message uses roughly 1,200 input tokens (system prompt + product context + a short history) and replies with about 120 Thai tokens. That's 6M input / 0.6M output tokens a month. At roughly ฿36 to the dollar (adjust for the rate when you read this), against API prices checked on 20 Aug 2026:

• •The hidden Thai surcharge — the same content in English would cut the bill by about 31% (Thai costs 1.46x the tokens). Sonnet 5 drops from ~฿972 to ~฿666: roughly ฿300 a month that buys the customer nothing
• •A free saving — many teams write the system prompt in English anyway, which leaves only the customer's message and the reply paying the Thai surcharge. Zero impact on the customer experience
• •Context pays it too — a 200K context holds about 500,000 Thai characters but about 910,000 English ones. Long Thai contracts fill the window almost twice as fast as people expect

Where LLMs still slip on Thai work — an opinion

In 2026 the risk is no longer "Thai output is bad" — current models write fluent Thai. The risk is a set of details only a Thai reader catches, and they slide through QA quietly:

• •Buddhist era vs Gregorian dates — the most common failure I hit. A document says "2569" and the model may silently treat it as a CE year. Any system touching Thai dates needs conversion and validation outside the model
• •Thai addresses — soi, khwaeng, khet, tambon and amphoe get swapped easily, and a postcode that doesn't match the district rarely triggers a warning. Validate against a real database before handing anything to a courier
• •Citing Thai law — models invent section numbers very fluently. For PDPA or contract work, let the LLM draft and summarise, but every section number must come from a document you supplied (that's what RAG is for)
• •Politeness particles drift — over a long conversation the model swaps khrap and kha and the brand voice falls apart. Pin it in the system prompt, then actually check the output
• •Wrong answers sound exactly like right ones — an LLM has no native "I don't know" signal. Anything where a mistake costs money needs a human or a rule at the end

LLM vs ChatGPT vs RAG vs AI Agent

These four get used interchangeably more than any other AI terms. Here is the split:

Summary: three things worth keeping

For a Thai business starting out, the order I recommend is chat app first → add RAG when answers must come from your documents → move to agents only when there's a repetitive process worth automating. Jumping straight to agents on day one is the most expensive and most failure-prone path I've watched people take. If you want help deciding where your business should start and what it realistically costs, talk to me about AI consulting.