AI models
The AI models, explained
Short answer
A handful of companies make the AI models most people use: Anthropic (Claude), OpenAI (GPT), Google (Gemini), Meta (Muse) and xAI (Grok), alongside open models from Mistral, DeepSeek and others. Each sells a range from small and cheap to large and capable, so the right choice depends on how hard the task is and how often you run it.
Model line-ups change every few weeks. This guide lists the main families and their newest widely reported models on the date shown. Each maker's own page is the authority.
Who makes which models?
Anthropic · Claude
Anthropic: Claude modelsClaude models are named by size and purpose. Haiku is small and fast, Sonnet is the balanced middle, and Opus is the starting point for most demanding coding and knowledge work. Fable and Mythos sit above them for the hardest, longest tasks.
| Model | What it is for | Released |
|---|---|---|
| Claude Fable 5.2 | Most capable generally available model, for long and difficult tasks | November 9, 2026 |
| Claude Mythos 5.2 | Restricted to verified life sciences and cybersecurity organizations | November 9, 2026 |
| Claude Opus 5.5 | Recommended starting point for most coding and knowledge work | — |
| Claude Sonnet 5.5 | Balanced speed, quality and cost | — |
| Claude Haiku 5.5 | Fastest and lowest cost | — |
OpenAI · GPT
OpenAI: modelsThe GPT-6 generation powers ChatGPT. Astra is the frontier model, with Sol and Luna as faster, cheaper tiers. OpenAI also makes image models under the GPT Image name.
| Model | What it is for | Released |
|---|---|---|
| GPT-6 Astra | Frontier model | September 3, 2026 |
| GPT-6 Sol | Faster, lower-cost everyday model | September 22, 2026 |
| GPT-6 Luna | Smallest and cheapest tier | September 22, 2026 |
| GPT Image 2.6 | Image generation | November 4, 2026 |
Google · Gemini
Google: Gemini modelsGemini comes in Pro, Flash and Flash-Lite tiers, trading capability for speed and cost. Gemini 4 Argon is the newest model and is in limited preview; Gemini 3.8 Flash is the newest one generally available.
| Model | What it is for | Released |
|---|---|---|
| Gemini 4 Argon | Newest model, in limited preview | October 2026 |
| Gemini 3.8 Flash | Newest generally available model, built for long coding and agent tasks | September 2, 2026 |
| Gemini 3.5 Flash-Lite | Low-cost, high-volume tier | — |
Meta · Muse
Meta AIMeta's Muse line pairs a large planning model with small, cheap models that carry out the steps. One member of the family, Glimmer, is open-weight, meaning anyone can download and run it.
| Model | What it is for | Released |
|---|---|---|
| Muse Wonder | Larger model for planning and reasoning | October 2026 |
| Muse Ember | Lightweight, low-cost model for agent tasks | November 9, 2026 |
| Muse Glimmer | Open-weight model | August 2026 |
| Muse Spark | First model in the Muse line | April 2026 |
xAI · Grok
xAI: modelsGrok is the model behind the assistant built into X.
| Model | What it is for | Released |
|---|---|---|
| Grok 4.7 | Current flagship | September 21, 2026 |
Mistral AI · Mistral
Mistral AI: modelsA French lab known for open-weight models that organizations can run on their own servers.
| Model | What it is for | Released |
|---|---|---|
| Mistral Large 4 | Open-weight flagship that handles text and images | October 29, 2026 |
| Mistral Small 4 | Small open model under the Apache 2.0 licence | — |
DeepSeek, Alibaba and Moonshot · Open models from China
DeepSeek: API docsChinese labs release some of the most widely used open and low-cost models.
| Model | What it is for | Released |
|---|---|---|
| DeepSeek V4 family | Low-cost models, including the V4.1 Flash variants | — |
| Qwen3.8-Max | Alibaba's flagship | August 3, 2026 |
| Kimi K3.1 Pro | Moonshot's open-weight model | November 2, 2026 |
Nvidia · GR00T (robotics)
NVIDIA: robotics announcements at CoRLNot every model is a chatbot. Nvidia's GR00T models control robots, turning camera input into physical actions.
| Model | What it is for | Released |
|---|---|---|
| GR00T 2 | Robot model trained largely on human video | November 9, 2026 |
How do I choose a model?
Start from the task, not the leaderboard. Three questions settle most choices.
- How hard is the task? Sorting emails or pulling figures from invoices needs only a small model. Long research, complex code or multi-step planning is where the large models earn their price.
- How often will it run? A task that runs thousands of times a day is dominated by cost per token, so a cheaper model can save more than a better one earns.
- Where can your data go? If information must stay on your own systems, look at open-weight models you can run yourself, or check the provider's data terms first.
New to the terms? How AI works explains parameters, tokens and training without the maths.
What are the latest model releases?
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic released two new top-tier models on November 9. Claude Fable 5.2 is generally available, including through AWS, Google Cloud and Microsoft Azure, at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the previous version. Claude Mythos 5.2 is limited to organizations in Anthropic's life sciences and cybersecurity verification programs. Anthropic says both models produce output up to 30% faster than their predecessors.
Why it matters: The most capable tier of AI is getting cheaper and faster, which lowers the cost of long, complex tasks for businesses.
Meta releases Muse Ember, a low-cost model built for AI agents
Meta released Muse Ember on November 9, describing it as the lightest and most cost-efficient model in its Muse line-up. It starts at $0.10 per million input tokens and is designed for high-volume agent work such as tool calls, routing, data extraction and parsing long documents. Meta pairs it with its larger Muse Wonder model: Wonder plans the work and Ember carries out the repetitive steps.
Why it matters: Cheap 'worker' models make it affordable to run AI agents across thousands of routine tasks.
Frequently asked questions
What is an AI model?
A model is the trained neural network itself: billions of numbers, called parameters, that were adjusted during training. An app such as ChatGPT or Claude is the product you use; the model is the engine inside it.
Why does each company have several models?
Bigger models are more capable but slower and more expensive to run. Companies sell a range so you can use a small, cheap model for simple, high-volume jobs and a large one for difficult work.
What does open-weight mean?
An open-weight model can be downloaded and run on your own computers, so your data never leaves your systems. Closed models are only available through the maker's app or API.
What is a token, and why are prices quoted per million tokens?
Models read and write text in small pieces called tokens, roughly three-quarters of a word each. Makers charge for the tokens a model reads (input) and writes (output), so prices are quoted per million of each.
Which model is the best?
It depends on the job. Rankings change every few weeks, and the most capable model is rarely the best value. For most everyday work a mid-tier model is enough; test two or three on your own task before choosing.
How this feed is made. The AI feed is researched and written each morning by Claude, an AI model made by Anthropic, and published automatically without human editing. Every story links to the reporting it is based on. AI can make mistakes, so check the source before relying on a detail, and treat nothing here as professional advice.