AI Model Families Explained
There are 412 AI models available today, but they all belong to a handful of model families. Understanding these families, their naming conventions, and their lineups makes the whole landscape easier to navigate.
How to Read Model Tiers
Every major provider offers models at multiple tiers. Flagship models are the most capable (and expensive). Balanced models offer the best quality-per-dollar. Fast models prioritize speed and low cost.
GPT
by OpenAI
The model family that started the AI revolution. Largest ecosystem and most integrations.
Key Strengths
Naming Convention
GPT-[generation].[version] with suffixes: -mini (small/cheap), -nano (smallest), no suffix = full size. The "o" series (o1, o3, o4-mini) are reasoning-focused models.
Evolution
GPT-3 (2020) -> GPT-3.5 (2022) -> GPT-4 (2023) -> GPT-4o (2024) -> GPT-4.1 (2025) -> GPT-5 (2025) -> o-series reasoning models
Current Lineup
Latest flagship. Top-tier reasoning and coding across all tasks.
Excellent balance of quality and cost. Great for production workloads.
Fast and cheap. Handles most tasks at a fraction of GPT-4.1 cost.
Dedicated reasoning model. Excels at math, logic, and multi-step problems.
Budget reasoning model. Good reasoning at lower cost than o3.
Claude
by Anthropic
Known for exceptional coding, careful reasoning, and natural-sounding writing. Strong safety focus.
Key Strengths
Naming Convention
Claude [generation] [tier]: Opus (flagship), Sonnet (balanced), Haiku (fast). Version numbers like 4.6 indicate the generation.
Evolution
Claude 1 (2023) -> Claude 2 (2023) -> Claude 3 (2024) -> Claude 3.5 Sonnet (2024) -> Claude 4 (2025) -> Claude 4.5 (2025) -> Claude 4.6 (2026)
Current Lineup
The most capable Claude. Leads on SWE-bench and complex reasoning.
Best value in the Claude family. Excellent coding at moderate cost.
Fastest Claude. Great for high-volume, latency-sensitive workloads.
Gemini
by Google
Massive context windows (up to 2M tokens), competitive pricing, and native multimodal capabilities.
Key Strengths
Naming Convention
Gemini [generation] [tier]: Pro (full power), Flash (fast/cheap). Ultra was the original flagship name but is now rarely used.
Evolution
Bard (2023) -> Gemini 1.0 (2023) -> Gemini 1.5 (2024) -> Gemini 2.0 (2025) -> Gemini 2.5 (2025) -> Gemini 3 (2026)
Current Lineup
1M token context. Strong reasoning with thinking mode. Excellent for research.
Extremely fast and cheap. Great default for production at scale.
Even faster and cheaper. For the highest-volume, cost-sensitive use cases.
Llama
by Meta
The most popular open-source model family. Free to use, modify, and deploy on your own infrastructure.
Key Strengths
Naming Convention
Llama [generation] [variant]: Scout (smaller), Maverick (larger). Parameter counts like 8B, 70B, 405B indicate model size.
Evolution
Llama (2023) -> Llama 2 (2023) -> Llama 3 (2024) -> Llama 3.1 (2024) -> Llama 3.3 (2024) -> Llama 4 (2025)
Current Lineup
Largest Llama 4 variant. Competes with proprietary models on many benchmarks.
Smaller, faster variant. Good balance for self-hosting on a single GPU server.
DeepSeek
by DeepSeek
Chinese AI lab delivering frontier performance at remarkably low prices. Open-weights for most models.
Key Strengths
Naming Convention
DeepSeek [series] [version]: V-series for general models, R-series for reasoning models.
Evolution
DeepSeek Coder (2023) -> DeepSeek V2 (2024) -> DeepSeek V3 (2024) -> DeepSeek R1 (2025) -> DeepSeek V3.1 (2025)
Current Lineup
Latest general model. Competitive with GPT-4.1 at a fraction of the price.
Reasoning-focused. Strong chain-of-thought for math and logic.
Mistral
by Mistral AI
European AI lab known for efficient architectures. Strong multilingual capabilities.
Key Strengths
Naming Convention
Named after weather patterns (Mistral, Pixtral) or size descriptors (Small, Medium, Large). "Codestral" is the coding-focused variant.
Evolution
Mistral 7B (2023) -> Mixtral 8x7B (2023) -> Mistral Medium (2024) -> Mistral Large (2024) -> Mistral Large 2 (2024) -> Mistral Small 3 (2025)
Current Lineup
Full-power model. Strong reasoning and multilingual capabilities.
Efficient, fast, and affordable. Great for production use.
Qwen
by Alibaba Cloud
Alibaba Cloud's open-source model family. Excellent multilingual performance, especially Chinese-English.
Key Strengths
Naming Convention
Qwen [generation] [size]: parameter count (7B, 32B, 72B, 235B). "A22B" suffix indicates active parameters in MoE models.
Evolution
Qwen (2023) -> Qwen 1.5 (2024) -> Qwen 2 (2024) -> Qwen 2.5 (2024) -> Qwen 3 (2025)
Current Lineup
Largest Qwen model. MoE architecture keeps inference costs low despite the size.
Sweet spot for self-hosting. Fits on a single high-end GPU.
Open Source vs. Closed Source
A key distinction that affects cost, privacy, and flexibility.
Closed Source (API-only)
GPT, Claude, Gemini
- + Typically highest performance
- + No infrastructure to manage
- + Regular updates and improvements
- - Pay per token, costs scale with usage
- - Data sent to third-party servers
- - Vendor lock-in risk
Open Source (Self-hostable)
Llama, DeepSeek, Mistral, Qwen
- + No per-token costs at scale
- + Full data privacy (runs on your servers)
- + Can fine-tune for your use case
- - Need GPU infrastructure
- - You manage updates and scaling
- - Slightly behind on the hardest benchmarks
Explore by Provider
See complete model lineups, pricing, and benchmark data for each provider.
There are 7 major model families in active development: GPT (OpenAI), Claude (Anthropic), Gemini (Google), Llama (Meta), DeepSeek, Mistral, and Qwen (Alibaba). Each family offers multiple variants at different capability and price tiers.
These are Anthropic's tier names for the Claude family. Opus is the most capable (and expensive), Sonnet is the balanced middle tier, and Haiku is the fastest and cheapest. Other providers use different naming: OpenAI uses suffixes (-mini, -nano), Google uses Flash/Pro/Ultra, and Meta uses Scout/Maverick.
Claude (Anthropic) currently leads in coding benchmarks like SWE-bench, followed closely by GPT (OpenAI). For open-source coding, DeepSeek and Llama perform well. The best choice depends on your budget and whether you need API access or self-hosting.
The gap is closing. Open-source models like Llama 4, DeepSeek V3, and Qwen 3 compete with or beat many closed-source models on standard benchmarks. Closed-source models (GPT-5, Claude Opus) still lead on the hardest tasks, but for most production use cases, open-source models are excellent.