Claude

Claude is Anthropic's family of large language models, named after Claude Shannon. Tier structure: Opus (largest, highest capability), Sonnet (balanced · the workhorse), Haiku (fastest, cheapest). Versioned numerically: Claude 3 (2024), Claude 3.5, Claude 4, Claude 4.5, Claude Mythos Preview (experimental 2026). Claude leads on coding benchmarks and excels at long-context comprehension and agentic tasks.

Claude models are decoder-only transformers trained with Constitutional AI · an alignment technique developed by Anthropic that uses AI-assisted self-critique during RLHF. Claude 4.5 Opus and Claude Sonnet 4 support Extended Thinking mode · hidden CoT tokens analogous to OpenAI o-series reasoning models. Claude models emphasize: long context (200K tokens default, 1M on Enterprise), tool use, agentic workflows, and coding performance (leading SWE-bench Verified scores). Anthropic serves Claude via direct API, AWS Bedrock, and Google Vertex AI · notable multi-cloud distribution.

Model sizes: undisclosed by Anthropic. Claude's post-training uses Constitutional AI (2022) for self-critique and RLHF (later RLAIF). Claude 3.5 Sonnet was the first model to cross 90% on SWE-bench Verified with the right agent scaffold. Extended Thinking is implemented as a budget parameter · models allocate up to 100K internal reasoning tokens per response, exposed transparently unlike OpenAI's hidden CoT. Context window handling uses high-quality long-context training (not just extension) · effective context preservation tops 90% at 200K, better than most peers.