What Is Claude, and What Is Anthropic?

Before you write a single line of code against the API, it helps to know what you're actually talking to — and who built it.

A large language model, demystified

☺ Like you're 10: It's the same trick as the autocomplete on your phone keyboard — guess the next word — just done by something that's read practically the whole internet and gotten very, very good at the guessing.

A large language model (LLM) is, at its core, a statistical model trained to do one thing: predict the next token — roughly a word or word-fragment — given everything that came before it. That's it. There's no hidden lookup table of facts and no separate reasoning engine bolted on underneath. What looks like knowledge or reasoning is really the pattern of relationships encoded across billions of parameters during training on enormous amounts of text.

Generating a full response is just this loop repeated: append the chosen token to the sequence, predict the next one again, and keep going until the model produces a stop signal. Trained across a huge and varied corpus — web text, books, code, and more — this simple mechanism, repeated at massive scale, is enough for a model to pick up grammar, facts, coding patterns, and problem-solving strategies, purely from having seen enormous numbers of examples of how text like it tends to continue.

Anthropic: safety as the starting point

☺ Like you're 10: Anthropic wrote Claude a rulebook before letting it talk to anyone, then had Claude practice grading its own homework against that rulebook instead of relying only on a teacher marking every single paper by hand.

Claude is built by Anthropic, which describes itself as "an AI safety and research company" and is structured as a Public Benefit Corporation with a stated purpose: the responsible development and maintenance of advanced AI for the long-term benefit of humanity. That framing isn't just marketing language — Anthropic treats AI safety as a systematic science, running frontier safety research, applying the resulting techniques to its own products, and publishing what it finds.

Two concrete commitments back this up: Claude's Constitution, a written set of principles the model is trained against, and a Responsible Scaling Policy, a framework for managing risk as model capabilities grow (covered in depth in this course's safety module). Anthropic frames its broader goal as trying to "ignite a race to the top on safety" across the AI industry, rather than treating safety as a constraint bolted onto a separately-optimized product.

One technique worth knowing the name of now is Constitutional AI. Instead of relying solely on large volumes of human-labeled examples of "good" and "bad" responses, Anthropic trains Claude against that written constitution in two stages. First, the model critiques and revises its own sample outputs against the constitution's principles. Then an AI, rather than a human, judges pairs of candidate responses for which one better satisfies those principles, and that AI-generated feedback trains a preference model used to further shape Claude's behavior. The goal is a model that's harmless without being evasive — one that explains its objections rather than just refusing.

The Claude model family

☺ Like you're 10: Think of it like a garage with a race car, a family sedan, and a scooter — same basic idea of getting somewhere, but you grab whichever one is actually sized to the trip you're taking.

Claude isn't one model — it's a family of models forming a ladder of capability, speed, and cost. Choosing a model is a real engineering decision: a more capable model costs more per token and may run slower, while a faster, cheaper model may be all a given task needs. The current generation, from most to least capable, is Claude Fable 5, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5, plus an invitation-only Claude Mythos 5.

{
  "model": "claude-sonnet-5",
  "max_tokens": 1024,
  "messages": [
    {"role": "user", "content": "Explain what a training cutoff is."}
  ]
}

That model string is exactly the identifier you'll pass in API requests once you get there. Here's how the current lineup compares:

A model's context window is how much text — measured in tokens, input plus output combined — it can consider in a single request. It's not the same thing as the model's trained-in knowledge; think of it as working memory for one conversation, not long-term memory. Every current Claude model also accepts image input alongside text (vision), in addition to producing multilingual text output, so the same model you use for code review can also read a screenshot or a scanned document.

Adaptive thinking vs. extended thinking

Some tasks benefit from Claude reasoning through a problem before answering, and there are two distinct mechanisms for this across the current lineup — they aren't interchangeable. Fable 5, Opus 5, and Sonnet 5 support adaptive thinking, where the model itself decides when and how much to reason based on the query; there's no manual switch to flip. Haiku 4.5 instead supports the older manual extended-thinking mode, turned on explicitly via a thinking parameter, but it doesn't support adaptive thinking. We'll dig into configuring and reading thinking output in a later module (see Extended Thinking) — for now, just know the two modes exist and which model tiers support which.

claude.ai vs. the Claude API and Developer Platform

☺ Like you're 10: claude.ai is like ordering a finished meal at a restaurant. The API hands you the same fresh ingredients so you can cook whatever dish your own app actually needs.

The same Claude models sit behind two very different products, and mixing them up trips up a lot of newcomers.

Training cutoffs: why Claude doesn't know today's news

A model's training cutoff is the point in time up to which the text it was trained on was collected. Whatever happened after that date simply isn't part of the data the model learned from — it has no way to know about it unless you tell it or give it a tool to look it up.

This is easy to conflate with the context window, but they're different things. The context window is what you feed Claude in a live request — you can absolutely paste in today's news article and ask Claude to summarize it, because that text is sitting right there in the conversation. The training cutoff is about the model's own baked-in knowledge — the facts and patterns it absorbed during training, frozen at that cutoff date, no matter how large its context window is.

This is exactly why agentic tool use matters, and why later modules spend real time on it: giving Claude a web search or web fetch tool lets it pull in current information at request time rather than relying solely on what it happened to learn during training.