AI in Plain Language

You do not need the maths. You need about ten words, because every decision you will make with an agent — which model, how much to paste in, when to start over, why the bill looks like that — is one of these ten words wearing a costume.

Read this once slowly. Then come back to it whenever a tool does something that makes no sense, because the explanation is almost always in here.

A model is a machine that finishes your sentence

A large language model is a program trained on an enormous amount of text to predict what comes next, one small piece at a time. Everything else — grammar, facts, code, the ability to follow an instruction — fell out of that single skill during training. It does not look answers up in a database; it generates them from patterns stored in billions of numbers called weights. The analogy that holds up: a very well-read person playing “finish my sentence” — fast, fluent, and occasionally confidently wrong.

That last part is not a defect you can patch out. It is the same mechanism working normally.

Everything you pay for is counted in tokens

A token is the unit the model actually reads and writes: a common word, a fragment of a word, or a punctuation mark. Text gets cut into these pieces from a fixed vocabulary that the tokenizer learned from its training data, so frequent strings become one token and rare strings get chopped into many. Every price, every limit and every “too long” error is measured in tokens, never in characters or words. Think of LEGO: some text snaps together from big pre-moulded pieces, other text has to be assembled from a pile of 1x1 bricks.

English needs roughly 1.2 to 1.8 tokens per word depending on the tokenizer. One thing worth knowing: Claude’s newer tokenizer produces about 30% more tokens for the same text than the old one, so 1M tokens is now around 555,000 words rather than the 750,000 the old rule of thumb promised (Anthropic model overview).

Kazakh costs several times more than English, and that changes how you work

This is the part that is specific to you. In the Sherkala paper’s Table 1, the Llama-3.1 tokenizer needs 4.73 tokens per Kazakh word, against 2.56 for Russian and 2.23 for Turkish; adding Kazakh tokens to the vocabulary cut that to 2.04 (arXiv). A separate 2026 measurement puts English at 1.23 tokens per word averaged across six public tokenizers (arXiv). Across 17 tokenizers, the gap between the cheapest and most expensive language reaches up to 15x (Petrov et al., NeurIPS 2023).

Three reasons stack up: Cyrillic letters take two bytes each in UTF-8, the Kazakh-specific letters (ә ғ қ ң ө ұ ү һ і) are rare in training data, and Kazakh is agglutinative — one word like “үйлеріміздегілерге” carries a whole English phrase. The working rule: Kazakh costs about 2x English on the best modern tokenizers and 3.5 to 4.7x on Llama-family ones. Same prompt, more money, faster-filling context window, slower generation.

The context window is a desk, not a filing cabinet

The context window is everything the model can see in one go: your instructions, the conversation so far, every file it read, every tool result, and the reply it is writing. Anything outside it does not exist for the model — there is no memory of your last session unless something put it back in. A 1M-token window is the normal ceiling now — Anthropic, OpenAI, Google and DeepSeek all reach it, and the price table at the end of this page carries the numbers with their sources. Not everyone does: Mistral stops at 256k, and Claude Haiku 4.5 at 200k. It is a desk: only the papers lying on it can be used, and when it overflows something gets filed away.

A bigger window does not mean a better answer

The practical version in Claude Code: /context shows you what is eating the window, /clear starts clean between unrelated tasks, and after two failed corrections on the same problem a fresh session with a better first prompt beats a long polluted one (Claude Code best practices).

Temperature decides how surprising the next word is allowed to be

Temperature is the randomness dial used when the model picks each next token. Near 0 it almost always takes the most likely option, so output is repeatable and dull; turn it up and less likely options get through, which buys you variety and risk in the same move. Low for code and data extraction, higher for brainstorming names. A chef at 0 follows the recipe exactly; the higher you turn the dial, the more they improvise. Note that some reasoning models set this themselves and ignore what you ask for.

Hallucination is the normal failure mode, not a rare bug

A hallucination is output that sounds right and is invented: a library function that does not exist, a citation with a real-looking author, a confident wrong price. It happens because the model generates plausible text rather than checking a source, and it almost never signals that it is unsure. Picture a student who did not study and still writes a fluent, confident exam answer. There are exactly two defences: grounding (give it the docs, let it search, let it run the code) and verification (tests, links, your own eyes). Anthropic’s guide now opens with the same idea — give the agent a way to verify its work, and ask for evidence rather than assertions (Claude Code best practices).

Before you answer, follow these rules:

1. List the files you actually opened, with line ranges. Do not answer from memory.
2. Answer only from what you read. If the answer is not in the repo, say
   "not in the repo" instead of guessing.
3. For every claim about how the code behaves, give me the command you ran
   and its real output. Evidence, not assertions.
4. If you are unsure, say which part you are unsure about and why.

My question: [your question here]

Tools are what let a model touch the real world

On its own a model can only produce text. Tool use (also called function calling) means you describe functions it is allowed to call — read a file, run a command, query an API — and the model replies with a structured request instead of prose. Your program executes that request and hands the result back, and the loop continues. It is a head chef who never touches the stove: precise orders out, tasted results back in.

An agent is a model running in a loop

An agent is the model plus tools plus a loop: look at the goal and the current state, choose an action, observe the result, repeat until done or until it needs you. Claude Code is exactly this — read files, edit, run the tests, read the error, fix, run again. The power comes from feedback, because the model is reacting to real results instead of guessing. The risk is that errors compound across turns, which is precisely why permission prompts, tests and git commits exist. A mechanic, not an oracle: try something, listen to the engine, adjust.

RAG is an open-book exam

RAG (retrieval-augmented generation) means searching your own documents first for passages relevant to the question, then pasting those passages into the prompt so the model answers from them. It gives the model fresh or private knowledge without retraining it, and lets it point at where an answer came from. The quality lives almost entirely in the search step: retrieve the wrong passage and you get a fluent wrong answer with a citation attached. Same student, same brain, but now the right page is open in front of them.

MCP is one connector instead of ten

The Model Context Protocol is an open standard for plugging external tools and data into AI apps: one server for GitHub, Figma or your database works in Claude Code, in Cursor, in other clients, without rewriting it each time. A server exposes tools (actions), resources (data) and prompts (templates); the AI app is the client. Think USB-C: one connector standard instead of a separate cable for every pair of devices. It started at Anthropic in November 2024, was adopted by OpenAI and Google DeepMind during 2025, and was donated in December 2025 to the Agentic AI Foundation under the Linux Foundation. The current spec version is 2026-07-28, which made the protocol stateless and deprecated the old HTTP+SSE transport (MCP specification).

Reasoning models think on scratch paper first

A reasoning model spends extra tokens working through a problem — planning, trying an approach, checking itself — before it writes the answer you see. That makes it much stronger at maths, debugging and multi-step planning, and slower and more expensive, because those thinking tokens are billed as output. In 2026 this is a dial rather than a separate product: Claude’s adaptive thinking with an effort setting, GPT’s reasoning levels, Gemini’s thinking budget. In Claude Code the dial is /effort, and the levels are low, medium, high (the default on most models), xhigh and max (model config docs).

What the models cost in September 2026

Prices below are US dollars per million tokens, input/output, checked 21 September 2026. They move, and some are explicitly temporary — Gemini 3.8 Flash’s price is promotional through 31 December 2026, after which it doubles. Open the pricing page the week you need a number; do not trust this table or any other one that is more than a month old.

Sources: Anthropic, OpenAI, Google, DeepSeek.

Abstract numbers mean nothing, so here is one job priced across all of them: roughly 100,000 input tokens plus 10,000 output tokens, which is about “read a mid-size codebase and write a feature”. Fable 5.1 costs about $1.50, Opus 5 about $0.75, Sonnet 5 about $0.30, Haiku 4.5 about $0.15, Gemini 3.8 Flash about $0.11, DeepSeek V4.1-Flash about $0.042 at peak rate, GPT-5.6 Luna about $0.032. That is roughly a 47x spread for one identical job. Gemini’s API still has a free tier, which makes it the realistic starting point if you have no card.