Lesson 4: Families of Models

The landscape

There are several major families of models available today. Each has different strengths, pricing, and context-window sizes. The choice depends on the use case.

Major families

Google Gemini

• Models: Gemini Pro, Gemini Ultra, Gemini Flash
• Strengths: Multimodal (text, images, video, audio), large context windows (up to 1M+ tokens), strong reasoning
• Best for: Tasks involving mixed media, long-document analysis

Anthropic Claude

• Models: Haiku (fast, cheap), Sonnet (balanced), Opus (most capable)
• Strengths: Strong instruction-following, long context windows (200K tokens), careful and safety-conscious outputs
• Best for: Code generation, analysis tasks, structured outputs

OpenAI GPT

• Models: GPT-4, GPT-4o, o1, o3
• Strengths: Broad general knowledge, strong coding ability, large ecosystem of tools and integrations
• Best for: General-purpose tasks, creative writing, code completion

Meta LLaMA

• Models: LLaMA 3, LLaMA 4 (open-weight)
• Strengths: Open-weight (can self-host), no API costs, customisable through fine-tuning
• Best for: Organisations that need data sovereignty or offline operation

Mistral

• Models: Mistral Large, Mistral Medium, Codestral
• Strengths: European-made, open-weight options, competitive performance at lower cost
• Best for: European regulatory compliance, cost-sensitive deployments

How to choose

Factor	Question to ask
Task type	Is this code, text, or multimodal?
Context size	How much input does the model need to process?
Cost	What is the per-token price? What is the expected volume?
Privacy	Can data leave your infrastructure?
Speed	Is latency critical for this use case?
Accuracy	How important is factual precision vs. creative output?

ReArch is model-agnostic — you can configure different models for different agents and switch between providers without changing your workflows.

---