The Model Landscape

Lecture 3.2

• Providers, brands, and models: the three-layer hierarchy (Anthropic → Claude → Sonnet)
• Model tiers across providers: fast/cheap, balanced, frontier — the consistent pattern
• The tradeoffs: capability, speed (latency), cost, context window size
• Frontier models: the most capable models available, and why "frontier" is a moving target
• Frontier hardware: hundreds of billions of parameters, GPU clusters, why these are cloud-only
• Billing: per-token pricing, input vs. output costs, ballpark estimates for common tasks
• Local models (Ollama, open-source): free inference, privacy, hardware requirements, quality gap
• Local hardware reality: ~1GB per billion parameters, what runs on a laptop vs. gaming GPU vs. multi-GPU
• Practical guidance: Sonnet as default, Haiku for iteration, Opus for hard problems

Additional Resources

• Lecture slides
• model_comparison.py (Python demo)