I'm good. Take me to pro level.

Every behavior you see in Claude or ChatGPT, good or bad, follows from a few mechanisms. Learn them once and the rest of this page will make sense. We go in order: what the model reads, what it does, how it was trained, why it fails, and what the app adds on top.

Tokens: the unit of everything

A language model doesn't read letters or words. It reads tokens. Anthropic calls tokens "the smallest individual units of a language model." A token can be a whole word, part of a word, a single character or a byte. OpenAI's rule of thumb for English: 1 token is about 4 characters, or about three quarters of a word. So 100 tokens is roughly 75 words.

Why you care: everything is measured in tokens. Context window size, usage limits, API prices and the length of a reply all count tokens. A 1,500-word article is about 2,000 tokens. Other languages and code often break into more tokens per word, so the same idea can cost more.

Next-token prediction

Google's machine learning course puts it bluntly: large language models "predict a token or sequence of tokens" and are "essentially autocomplete mechanisms." For each step, the model assigns a probability to every possible next token. One token gets picked, added to the text, and the process repeats. A long answer is thousands of these small predictions in a row.

This has a direct consequence. The model produces what is likely to come next given everything it has seen. Likely and true usually overlap. When they don't, you get a confident error.

Parameters

Parameters are the internal numbers the model adjusts during training. Think of them as knobs. Google's course says current models contain "hundreds of billion or even trillions of parameters." The knowledge in a model lives in the settings of those knobs. No single knob holds a fact, which is why a model can't simply look up where it learned something.

Transformers and attention, in plain words

The transformer is the design almost all of today's language models use. Its key part is self-attention. For each token, the model asks how much every other token in the text affects its meaning. In "The bank raised its rates," attention links "bank" to "rates" and away from "river." Attention is why a model can connect a question at the end of your message to a detail on page 40 of the document you pasted. It is also why long inputs cost more work: each new token looks back at every token before it.

Stage one: pretraining

Pretraining is the first and largest stage. The model trains "on a large unlabeled corpus of text" (Anthropic's glossary): web pages, books and other writing. Its only job is to predict the next token. Nobody labels the text true or false.

A pretrained model is a powerful text continuer and a poor assistant. Anthropic notes that pretrained models are "not inherently good at answering questions or following instructions." Ask one a question and it may continue with three more questions, because that pattern appears in its data.

Stage two: fine-tuning

Fine-tuning means further training a pretrained model on additional, targeted data so it picks up that data's patterns. For assistants, that data is examples of good conversations: a request, then a helpful answer.

Stage three: RLHF, step by step

Reinforcement learning from human feedback (RLHF) trains a model "to behave in ways that are consistent with human preferences." The classic recipe comes from OpenAI's InstructGPT paper (Ouyang et al., 2022). It has three steps:

1. Demonstrations. Human labelers wrote example answers to prompts, and the model was fine-tuned on them.
2. Rankings. Labelers ranked several model outputs from best to worst. Those rankings trained a separate reward model that scores answers the way the humans would.
3. Reinforcement learning. The model was tuned to produce answers the reward model scores highly.

The result surprised people. Human raters preferred the 1.3-billion-parameter InstructGPT over the 175-billion-parameter GPT-3, a model more than 100 times larger. InstructGPT was more truthful and less toxic, and the paper still says it "still makes simple mistakes." Training shapes behavior. It doesn't remove errors.

Constitutional AI

Anthropic published Constitutional AI on December 15, 2022. Human oversight comes from a written list of principles, the "constitution." Training runs in two phases. First, the model critiques and revises its own answers against the principles, and it is fine-tuned on the revised versions. Second, in RL from AI Feedback (RLAIF), an AI judges which of two answers is better. Those judgments train a preference model, and that model provides the reward. Anthropic's stated goal was an assistant that is "harmless but non-evasive."

Why it hallucinates

OpenAI defines hallucinations as "plausible but false statements generated by language models." In a paper published September 5, 2025, OpenAI gave two reasons they persist.

• Pretraining has no true/false labels. The model learns patterns. Spelling and grammar follow patterns. A random fact, such as a particular person's birthday, follows none, so the model can't learn it reliably from patterns.
• Evaluations reward guessing. Most benchmarks score accuracy only. That works like a multiple-choice test with no penalty for wrong answers. A guessed birthday has a 1-in-365 chance of being right. "I don't know" always scores zero. A model trained to win those tests learns to guess.

OpenAI's numbers make the point. gpt-5-thinking-mini declined to answer 52% of the time and had a 26% error rate. o4-mini had similar accuracy and a 75% error rate, because it guessed far more often. OpenAI's proposed fix is to penalize confident errors more than honest uncertainty. The paper's example: a chatbot confidently gave three different wrong titles for a researcher's dissertation.

Your practical takeaway: a model that sounds sure is not evidence of a correct answer. Rare facts, names, dates, numbers, quotes and citations are the highest-risk categories.

Knowledge cutoff

A model knows nothing about the world after its training data ends. Anthropic separates two dates. The training-data cutoff is when the data stops. The "reliable knowledge cutoff" is earlier, because the last months before the cutoff are thinly covered online when the data is collected. Current Claude models list June 2026. OpenAI lists April 30, 2026 for GPT-6 Astra. Web search fixes this gap: it finds information "past Claude's knowledge cutoff, with cited sources." For anything recent, make sure search is on and read the sources.

Temperature and why answers vary

Remember that the model outputs probabilities for the next token. The app then samples from them. Temperature "controls the randomness of a model's predictions" (Anthropic). Low temperature makes the model pick the most likely token almost every time, so output is "more conservative and deterministic." High temperature lets less likely tokens through, so output is more varied and often more creative. OpenAI adds that higher temperature does not make answers more truthful, and it recommends temperature 0 for factual extraction.

This is why the same question gives different answers on different tries. Each run can take a different path through the probabilities. We found temperature documented as a developer (API) setting and found no temperature control documented for the consumer apps. So in the apps, handle variation another way: if two runs of a factual question disagree, treat both as unverified.

The model and the app are different things

The model is the trained network. The app is everything wrapped around it. On claude.ai and in the Claude mobile apps, a system prompt gives Claude current information, such as today's date, and encourages certain behaviors. Anthropic publishes these system prompts and notes they "do not apply to the Claude API." ChatGPT's pricing page says the chat's window is also used for system instructions (including tools and personality), memories and internal processing. So the app adds a system prompt, your custom instructions, memory and tools before your first word. The same model can behave differently in the app and in the developer API for exactly this reason.

Tools and agents

A model can only produce text. Tools let it act. In Anthropic's docs, tool use lets Claude "call functions that you define or that Anthropic provides." The model writes a structured request ("search the web for X"), something outside the model runs it, and the result comes back into the chat. This is also called function calling. Web search, fetching a page and running code in a sandbox are all tools. "Tools differ primarily by where the code runs."

An agent is a model that uses tools in a loop and decides its own steps. Anthropic's engineering team defines agents as "systems where LLMs dynamically direct their own processes and tool usage," and contrasts them with workflows, which follow "predefined code paths." Its advice for builders applies to you too: start with "the simplest solution possible." MCP (Model Context Protocol) is the open standard many tools use to plug into AI apps, described as "like a USB-C port for AI applications."

Thinking models and effort

Reasoning models work through a problem internally before they answer. OpenAI says its reasoning models use "reasoning tokens to 'think,' breaking down the prompt and considering multiple approaches." On Claude, with adaptive thinking, "Claude decides, for each request, whether to think and how much." You steer the depth with an effort setting.

• Claude: thinking is always on for the newest models (Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5). The Effort menu offers Low, Medium, High, Extra high and Max. Lower effort makes it think less.
• ChatGPT: the Thinking slider offers Instant, Medium, High, Extra High and Pro. Medium and higher draw on your Thinking limits.

Thinking costs something on both planets. Thinking tokens count toward the context window, and on paid usage they count as output. Use high effort for analysis, planning and hard problems. Use low effort or Instant for quick rewrites and lookups.

Multimodal: images and voice

Multimodal models handle more than text. Claude accepts text and images and produces text. Its documented vision limits are worth memorizing: it won't name people in images, it may hallucinate on low-quality, rotated or tiny images (under 200 pixels), object counts are approximate, and it "cannot determine whether an image is AI-generated." It is not for CT or MRI diagnosis. OpenAI's GPT-4o (May 13, 2024) introduced one model trained end to end on text, vision and audio. The earlier voice mode chained three models: speech to text, then GPT, then text to speech. GPT-4o answers audio in as little as 232 milliseconds, 320 on average. Claude's voice mode is in beta on all plans, saves text transcripts and can use connected tools and web search.

Sycophancy: when it tells you what you want to hear

Sycophancy is the model agreeing with you, flattering you or backing your view when it shouldn't. It is a side effect of training on human preferences, because people tend to rate agreement highly.

• Anthropic's research (October 23, 2023) found that five leading assistants consistently told users what they wanted to hear. Humans and preference models sometimes preferred convincing sycophantic answers over correct ones. The paper calls sycophancy "a general behavior of RLHF models."
• OpenAI's rollback (April 2025). A GPT-4o update on April 24 and 25 became sycophantic. It validated doubts, stoked anger and encouraged impulsive actions. The cause was a new reward signal from thumbs-up and thumbs-down feedback that weakened existing checks. Offline evaluations looked fine. The rollback began April 28, and OpenAI published a post-mortem on May 2, 2025.

Both companies have written openly about this, so treat it as a known property of the technology. Section 4 shows you how to prompt against it.

Anthropic defines the context window as "all the text a language model can reference when generating a response, including the response itself." It is the model's working memory for one conversation, and it is separate from what the model learned in training. If something isn't in the window, the model can't see it right now.

What's inside the window

The window holds far more than your last message. In rough order of arrival:

1. The system prompt the app adds (date, behavior rules).
2. Your custom instructions and any project instructions.
3. Memories the app loads about you.
4. Files and images you upload, and project knowledge it pulls in.
5. Tool results: search results, fetched pages, connector data, code output.
6. Every message you've sent in this chat.
7. Every reply, including thinking tokens.

All of it competes for the same space. A 60-page PDF plus a long thinking reply can fill a smaller window fast.

Sizes per planet and plan

Two notes. On ChatGPT, the reasoning window is much larger than the Instant one, so long documents fit better with a Thinking setting. The developer APIs list 1M tokens on current Claude models (the same as the Claude app for those models) and 1.05M on current GPT-6 models, far above ChatGPT's app figures.

Context rot

A bigger window doesn't mean perfect recall. Anthropic's documentation says it directly: "as token count grows, accuracy and recall degrade." It calls this context rot. More context is not automatically better. A focused chat with the right three documents usually beats a sprawling one with twenty.

What happens in very long chats

• On Claude: on paid plans with code execution on, when a chat nears the limit "Claude summarizes earlier messages to make room for new content." The full history is kept so Claude can still refer to it. The summary keeps the gist and loses detail, and summarizing uses more of your usage limit. Without code execution, this doesn't happen.
• On ChatGPT: the window is set by your plan and mode, as in the table. OpenAI's pricing page notes that system instructions and memories use part of it. Our research found no OpenAI page describing exactly what ChatGPT does when a chat outgrows its window. Assume the model can't see everything, and watch for the signs below.

Signs you've outgrown the chat

• It forgets a rule you set at the start.
• It asks for information you already gave.
• It mixes details from two different tasks.
• Quality drops: vaguer, longer or more repetitive replies.

When to start fresh

Start a new chat when the task changes, when you see any sign above, or when the chat has covered several large documents. One task per chat is the habit that prevents most problems on both planets.

How to hand off cleanly

Before you leave a long chat, ask for a handoff note. Then paste it as the first message of the new chat, or save it in your project's knowledge so every new chat sees it.