Claude (Anthropic)

Claude is a family of large language models (LLMs) developed by Anthropic, designed for safe and helpful text generation. Operators encounter Claude primarily through Anthropic's API (claude.ai) or the console, as it is not open-weight and cannot be run locally. Claude models (e.g., Claude 3.5 Sonnet, Claude 3 Opus) are accessed via HTTP requests, returning generated text. For local-AI operators, Claude represents a closed-source alternative to open-weight models like Llama or Mistral, with different pricing, latency, and safety alignment.

Claude models are built using transformer architectures with a focus on constitutional AI—training to follow a set of principles rather than relying solely on human feedback. The family includes tiers: Haiku (fast/cheap), Sonnet (balanced), and Opus (most capable). Operators using Claude via API must manage API keys, rate limits, and costs (per-token pricing). Unlike local models, Claude offers no control over hardware, no offline use, and no ability to fine-tune or quantize. Latency varies by model tier and request size, typically 1-5 seconds for short responses. For local-AI operators, Claude is a reference point for quality benchmarks but not a deployable runtime component.

An operator comparing local vs. cloud models might run Llama 3.1 8B locally at 40 tok/s on an RTX 4090, while Claude 3.5 Sonnet via API might respond at ~100 tok/s but with network latency (200ms) and per-token cost ($3/M input tokens). The trade-off: local offers free inference after hardware cost, privacy, and no rate limits; Claude offers higher quality but recurring costs and data sent to Anthropic.

In a workflow, an operator might use Claude for complex reasoning tasks via curl: curl https://api.anthropic.com/v1/messages -H "x-api-key: $ANTHROPIC_API_KEY" -d '{"model":"claude-3-5-sonnet-20241022","messages":[{"role":"user","content":"Explain quantum entanglement"}]}'. The response is JSON with generated text. For local fallback, the operator might switch to Ollama with ollama run llama3.1 when API costs exceed budget.