The Major Model Families
The AI model landscape in 2025–2026 is dominated by a handful of major families, each with distinct strengths, licensing models, and use-case sweet spots. This page covers each family in enough depth to make informed selection decisions.
Closed API = pay-per-token, zero ops. Open-weight = fixed infra cost, full control. Most production systems use both. (June 2026)
OpenAI - GPT and o-Series
OpenAI maintains two parallel model lines for different needs:
GPT Series
| Model | Context | Strengths | Best for |
|---|---|---|---|
| GPT-4o | 128K | Multimodal, fast, broad capability | Cost-efficient general tasks; vision |
| GPT-4o mini | 128K | Fast, very cheap | High-volume simple tasks |
| GPT-5 | 400K | High general capability, reduced hallucination | Professional knowledge work, hardest general tasks |
| GPT-5.5 (Apr 2026) | 400K | Multi-step agentic tasks, coding, tool use; more intelligent and token-efficient than GPT-5 | Current OpenAI flagship - autonomous task completion |
| GPT-5.5 Pro (Apr 2026) | 400K | Research-grade complexity, maximum OpenAI capability | Hardest reasoning tasks; $30/$180 per 1M tokens |
o-Series (Reasoning)
OpenAI's separate reasoning-focused family. Spends additional compute "thinking" before answering. Dramatically better at maths, formal logic, and complex code. See the Reasoning Models section for full detail.
- o3 - Full reasoning with native tool use; best quality on hardest problems
- o4-mini - Cost-efficient reasoning; on benchmarks often matches o3 at 1/9th the cost
OpenAI's core strength: Largest developer ecosystem, broadest tool integration (Assistants API, function calling, fine-tuning), most mature production infrastructure.
Anthropic - Claude Family
Anthropic's Claude models are known for instruction-following quality, long-context accuracy, coding reliability, and safety-focused behaviour. The naming convention:Haiku (fast/cheap) → Sonnet (balanced) →Opus (most capable).
| Model | Context | Strengths |
|---|---|---|
| Claude Haiku 4.5 | 200K | Fastest Claude, very cheap, good for routing and classification |
| Claude Sonnet 4.6 | 200K | Best all-round value: coding, analysis, writing, agentic tasks |
| Claude Opus 4.8 (May 2026) | 200K | #1 on Artificial Analysis Intelligence Index (score 61.4); 84% on browser-agent benchmark; only model to complete every Super-Agent case end-to-end; dynamic workflows for large-scale tasks |
Claude's core strengths: The current frontier for agentic and computer-use tasks - Opus 4.8 outperforms GPT-5.5 on browser-agent evals. Extremely long and accurate context handling (200K native), strong coding reliability, and the best model for autonomous multi-step agents that need to operate a browser or desktop. Claude Code (Anthropic's agentic coding tool) is built on Claude Sonnet/Opus.
Google - Gemini Family
Google's Gemini family is defined by its multimodal capability and extreme context lengths.
| Model | Context | Strengths |
|---|---|---|
| Gemini Flash 2.5 | 1M | Very fast, 1M context, multimodal, cheap |
| Gemini 3.5 Flash (GA May 2026) | 1M | 284 tokens/sec; Intelligence Index score 55; strong coding and agents; fast-mode default for Google-stack builders |
| Gemini 3.5 Pro (Jun 2026) | 2M | 2M token context; Deep Think reasoning mode; frontier multimodal; ~$15/$60 per 1M tokens |
Gemini's core strengths: The 2M token context window on Gemini 3.5 Pro is the largest in production - making whole-codebase analysis of even large monorepos practical. Gemini 3.5 Flash at 284 tokens/second is the fastest frontier-class model for throughput-sensitive agentic pipelines. Deep Google Workspace integration, and now powering Siri AI in iOS 27.
Meta - LLaMA Family (Open-Weight)
LLaMA is the world's most widely used open-weight model family. Meta releases model weights freely for research and commercial use (check specific version licences).
| Model | Parameters | Key features |
|---|---|---|
| Llama 3.1 8B | 8B | Runs on consumer GPU; good general reasoning |
| Llama 3.1 70B | 70B | Near-frontier quality; runs on 2× consumer GPUs |
| Llama 3.1 405B | 405B | Strong dense model; requires data centre GPU; superseded by Llama 4 for most uses |
| Llama 4 Scout | 109B MoE (17B active) | Best open-weight quality in class (Apr 2026); 10M context; multimodal vision; fits one H100 |
| Llama 4 Maverick | 400B MoE (17B active) | Matches GPT-4o on benchmarks; multimodal; inference cost of a 17B model; production API use |
LLaMA's core strength: No per-token cost, full data control, fine-tuning freedom. The massive open-source ecosystem (Ollama, llama.cpp, vLLM) makes deployment straightforward. Llama 4 achieves 85–86% MMLU-Pro - matching or approaching proprietary frontier models on many benchmarks.
Mistral - Compact European Models
Mistral AI (Paris) makes high-efficiency open-weight models with a strong commercial presence in Europe:
- Mistral 7B / Nemo (12B) - Punches above its weight for size; particularly strong on coding and instruction-following
- Mixtral 8x7B / 8x22B - Mixture-of-Experts models; only a subset of experts active per token, making them cost-efficient
- Mistral Large 3 - 675B total, 41B active (sparse MoE); closed API; competitive with frontier models on European-language tasks; Apache 2.0 for smaller models
Mistral's niche: European regulatory comfort (French company, EU-hosted options), strong multilingual European languages, and the most efficient open models for their quality tier.
DeepSeek - Chinese Open-Weight Leader
DeepSeek has produced the most impactful open-weight releases of 2024–2025:
- DeepSeek-V3 / V3.2 - Non-reasoning model; 671B MoE, 37B active; strong on coding and general tasks; very cheap API ($0.27/1M input tokens for V3)
- DeepSeek-R1 - Open-weight reasoning model that matches o1; full reasoning trace visible; distilled versions available for local deployment
DeepSeek's impact: R1's January 2025 release caused a "global cost reset" - demonstrating frontier reasoning capability is achievable without massive closed-source infrastructure. However: data privacy and safety alignment considerations apply (Chinese training and governance).
Microsoft - Phi and MAI Models
Microsoft now has two distinct model lines: Phi (small/edge) and MAI (frontier, built in-house without OpenAI distillation):
- Phi-3 Mini (3.8B) - Runs on phones; strong reasoning per parameter
- Phi-4 (14B) - Improved quality; strong on STEM tasks; can run on laptop GPU
- MAI-Code-1-Flash (5B active) - Agentic coding specialist; deep GitHub Copilot integration; available to all Copilot plans (Jun 2026)
- MAI-Thinking-1 (35B active, 256K context) - Microsoft's first in-house reasoning model; matches Claude Sonnet 4.6 in blind evals; private preview on Azure AI Foundry (Jun 2026)
Microsoft's shift: The MAI model line marks a strategic move away from pure OpenAI dependency. MAI-Code-1-Flash becoming the GitHub Copilot backbone means Microsoft now controls the model powering the most widely used AI coding assistant, with pricing and roadmap independent of OpenAI.
Alibaba - Qwen3 (Multilingual Open Models)
Qwen3 (released May 2026, Apache 2.0) is the most capable open-weight family for multilingual tasks and coding - 29 languages, with a thinking-mode toggle per request:
- Qwen3 0.6B–30B - Full range for edge to server; thinking mode on/off per request
- Qwen3 72B - Dense model; strong instruction-following, coding, and math
- Qwen3 235B MoE - Total params / ~22B active; frontier-class open model
- Qwen3 Coder 480B MoE - 262K context; targets frontier coding models at open-weight prices; available free on OpenRouter
Qwen3's key innovation: The thinking-mode toggle - use fast mode for simple tasks, reasoning mode for complex ones, in the same model per request. Previously you needed separate models for this. Qwen3 is also the base architecture that DeepSeek R1 distillation builds on.
Reading Model Naming Conventions
Common patterns you'll encounter across families:
- Size suffix (7B, 70B, 405B) - Billions of parameters. Larger = more capable but more compute to run
- Instruct / Chat / Base - "Instruct" or "Chat" = fine-tuned to follow instructions. "Base" = raw pre-trained weights (not for end users)
- Q4, Q8 (quantisation) - Weight precision reduced to save memory. Q4 = 4-bit; sacrifices some quality for much smaller file size
- GGUF - File format for local inference via llama.cpp/Ollama
- MoE (Mixture of Experts) - Total params / Active params notation (e.g., 8x7B = 8 experts of 7B each; only 2 active per token = 14B active)