Claude, overview

Claude is Anthropic's family of large language models. For autonomous AI, Claude matters because it was designed with agents in mind, strong tool use, long context, and built-in safety tuning.

The families

Claude comes in three tiers at any given generation:

• Opus, the most capable. Highest reasoning, best on complex tasks. Most expensive.
• Sonnet, the balanced choice. Strong across most tasks, reasonable price.
• Haiku, fast and cheap. Good for high-volume, simpler work, or as a "worker" model in multi-agent setups.

Pick based on task complexity and cost tolerance. Most agents work well on Sonnet. Spin up Opus for hard reasoning, drop to Haiku for parallel, cheap worker tasks.

Why Claude for agents

• Tool use is first-class. Structured tool-calling with retries, JSON schemas, and parallel calls baked into the API.
• Long context window. Up to 1M tokens on some models, enough for large codebases or multi-document RAG.
• Extended thinking. The model can reason internally before answering. Useful for hard planning tasks. See Extended thinking.
• Prompt caching. Materially cheaper for repeated-context workflows like agents.
• Safety tuning. Trained to refuse obviously bad requests and to be careful with dangerous actions.

How it compares

Honest take: GPT-4 and Gemini are also strong. For pure language tasks they're comparable. For agent work, Claude has a real edge on:

• Multi-turn tool-use stability (fewer dropped calls)
• Following complex system prompts
• Refusing obviously bad tool calls when given the right framing

For image reasoning or code, the race is closer. Vendor lock-in is a real risk, prefer abstractions that let you swap models.

API vs. Claude.ai vs. Claude Code

Three surfaces matter:

• Claude API. Direct programmatic access. Use for building your own agents.
• Claude.ai. The chat interface. Good for exploration, not for production agents.
• Claude Code. The CLI agent harness. Anthropic's opinionated wrapper around the API with permissions, hooks, skills, and MCP support. See Claude Code overview.

Pricing model (directional)

Priced per-token, input and output separately. Output tokens cost 5× input. Prompt caching reduces input cost by ~90% on cached portions. For agent workflows with stable system prompts, caching matters a lot. See Prompt caching.