Claude GPT Gemini Llama and Mistral
Models, Families and Providers
A model is one trained network with a fixed set of weights, such as claude-sonnet-5-5. A model family is a series of models from one company that share a design and are released in versions and sizes (Claude, GPT, Gemini, Llama). A provider is whoever runs the model and gives you access to it. Sometimes that is the company that built it (Anthropic for Claude), and sometimes a cloud platform that hosts it (AWS Bedrock, Google Vertex AI, Microsoft Foundry) or a hosting service for open models.
All of these families are built on the same basic idea you learned in How LLMs Work: transformer networks trained on next-token prediction, then tuned to follow instructions. They differ in training data, size, the quality of their tuning, the extra abilities they support (images, PDFs, tool use, long context), price, and the licence under which you may use them. No single family is best at everything, and rankings change every few months, so learn to compare rather than memorise a winner.
The Main Families
| Family | Company | Type | Access | Known for |
|---|---|---|---|---|
| Claude (Opus, Sonnet, Haiku) | Anthropic | Closed | Claude API, AWS Bedrock, Google Vertex AI, Microsoft Foundry | Coding, agents, long documents, careful answers |
| GPT | OpenAI | Closed | OpenAI API, Azure OpenAI | ChatGPT, broad ecosystem |
| Gemini | Closed (Gemma is open) | Gemini API, Vertex AI | Multimodal, Google integration | |
| Llama | Meta | Open weights | Download, Ollama, Hugging Face, cloud hosts | Self-hosting, fine-tuning |
| Mistral / Mixtral | Mistral AI | Open and closed | Mistral API, download | Efficient European models |
| Others | DeepSeek, Qwen (Alibaba), Sarvam, Krutrim | Mostly open | Download, APIs | Reasoning, Indian languages |
Claude Model Tiers
Most families come in tiers: a large, most capable model; a balanced middle model; and a small, fast, cheap model. The tiers trade quality against speed and price. Larger models handle harder reasoning and long multi-step tasks better; smaller ones answer faster and cost a fraction as much, which matters when you make millions of calls.
| Model ID | Tier | Price in/out per MTok | When |
|---|---|---|---|
claude-opus-5-5 | Most capable Opus | $4 / $20 | Default for hard tasks, agents, coding |
claude-sonnet-5-5 | Balanced | $2 / $10 | High-volume production chat and RAG |
claude-haiku-5-5 | Fast and cheap | $0.10 / $0.50 | Classification, extraction, routing |
Note: Use the exact model IDs above in code. All three have a 1M-token context window. Prices change; check the pricing page.
Misconceptions and Summary
Note: "ChatGPT is the model." ChatGPT and claude.ai are chat apps; developers use the models through an API, without the app's extra features or memory. "Benchmarks tell you which model to use." Public benchmarks are useful signals, but your own task, language and data can rank models differently. "Model names are stable." Providers release new versions and retire old ones, so keep the model ID in configuration, not scattered through your code.
- A model is one set of weights; a family is a series; a provider is who serves it.
- All major families are transformer LLMs; they differ in data, tuning, features, price and licence.
- Families come in tiers that trade capability for speed and cost.
- Compare models on your own task instead of trusting a single ranking.