What is Claude?

What is Claude?

Claude is the AI assistant built by Anthropic - a family of large language models designed for conversational reasoning, long-form writing, code generation, document analysis, and tool use. You talk to it the same way you would talk to a thoughtful collaborator: through natural language, in plain text, with as much or as little context as you want to give.

Under the hood, Claude is a transformer-based large language model in the same broad architecture family as GPT, Gemini, and Llama. What sets it apart is less the architecture and more the priorities that shaped it: helpfulness, honesty, and harmlessness are the explicit design goals, and the training methods Anthropic uses (Constitutional AI, debated in Section 06) are built around making that triple stick under pressure.

Talk to Claude, an AI assistant from Anthropic.- claude.ai

Who built it

Anthropic is the AI safety company that builds and trains Claude. It was founded in 2021 by Dario and Daniela Amodei along with several colleagues from OpenAI's research staff. The company's stated mission is to research and build safe, beneficial AI systems - and Claude is the most visible result of that work.

The split from OpenAI is part of the story. Anthropic's founders left because they believed AI safety needed to be a research priority at every layer of the stack - data, training, model behavior, deployment - rather than a constraint applied at the end. That belief is what makes Claude's design feel different from a model that was first optimized for capability and then bolted-with safety filters afterward.

The Claude model family

"Claude" is not one model. It is a family, with multiple tiers released in parallel that trade off capability against speed and cost. The tier names follow a consistent pattern across generations:

Each generation of Claude (3, 3.5, 4, and onward) ships with new versions of these tiers. Picking the right tier is almost always more impactful than picking the right generation - a fast tier on a recent generation will usually beat a slow tier on an older one for the same dollar.

What Claude is good at

Across benchmarks and across the actual things people use it for day to day, Claude clusters in five areas:

• Conversational reasoning. Multi-turn discussions that build on prior context. Asking follow-up questions, exploring trade-offs, working through a problem out loud.
• Long-form writing. Drafting, editing, restructuring, and adapting written material - reports, documentation, emails, narrative. Strong at maintaining tone and structure across long outputs.
• Coding. Generating, reviewing, and refactoring code in most languages. Strong at structured outputs, careful diffs, and understanding what a codebase already does before suggesting changes.
• Document analysis. Reading long PDFs, transcripts, and large amounts of context, then synthesizing or answering questions across all of it. The Claude family ships with long context windows (200K+ tokens in current models).
• Tool use and agents. Calling tools - APIs, file systems, browsers - as part of completing a task. The basis for Claude Code (the CLI tool) and for the broader agent-skill model Anthropic publishes.

What Claude is less suited for: anything that demands exact factual recall about niche topics without retrieval, real-time information about current events without browsing, or tasks where an LLM is the wrong tool entirely (cryptographic precision, financial settlement, hard math without code).

Where you can use Claude

Claude is available through several surfaces, each suited to a different kind of work:

Constitutional AI and safety

Most modern LLMs are aligned through Reinforcement Learning from Human Feedback (RLHF) - humans rate model outputs, the model is fine-tuned to produce outputs humans rate well. Claude is too, but with an additional Anthropic technique called Constitutional AI (CAI).

In CAI, the model is given a short written "constitution" - a set of principles describing how a helpful, honest, harmless assistant should behave - and is trained to critique and revise its own outputs against that constitution. The constitution is public. Anthropic has published the principles it uses and writes about how they evolve.

What this gets you in practice:

• More predictable behavior under pressure. Claude is reluctant to be talked out of its principles by clever framing or roleplay - the constitution gives it something to reason against instead of just trained-in reflexes.
• Transparent guardrails. The values shaping Claude's behavior are written down. You can read what they are. This is unusual.
• An honest "I don't know." Claude is more likely than some peer models to say it isn't sure, decline a request it can't safely complete, or push back on an instruction it thinks is wrong.

None of this is perfect - no current alignment technique is. But the design intent is visible, which makes it easier to reason about how Claude will behave at scale.

When to reach for Claude

Claude is the right call when the task involves real reasoning, careful writing, long context, or coding - and when you want a model whose alignment behavior is predictable and documented. Pair it with TypeScript for API integrations, with retrieval and tool use for grounded answers, and with Claude Code for code-heavy workflows.

Reach for something else when the task is fundamentally about generating images or video (different model families do that better), when you need real-time web access without configuration, or when budget is so tight that even Haiku is too expensive for the volume you need.

For everything in the broad middle - drafting, reasoning, coding, analysis, support, agents - Claude is the model most teams reach for, and the model most of this article's resources are about.