Start here4 min read
AI Model Guide: Claude, GPT, Gemini, Grok and More
What each name and version means: Fable, Opus, Sonnet, Astra, Sol, Luna, Gemini, Grok, Muse Spark and DeepSeek. What each is for and what it costs.
A new model with a new name comes out every few weeks, and it is easy to get lost among Opus, Sol, Argon and Spark. The good news is that nearly every company follows the same logic: a family in several sizes and a version number that goes up with each improvement. Understand that and you understand any announcement. This guide is current as of October 2026; exact prices are in the model comparison.
How to read a model name
- The family name tells you the size. The big one is the most capable, slowest and most expensive; the small one is the fastest and cheapest. At Anthropic, from largest: Fable, Opus, Sonnet and Haiku. At OpenAI: Astra, Sol and Luna. At Google: Pro and Flash.
- The number tells you the generation. Opus 5.5 improves on Opus 5 within the same generation; a jump to 6 would be a new generation. A “.1” is usually a refinement: GPT-6.1 Sol improves GPT-6 Sol.
- “Preview” or “beta” means provisional. Behaviour or price may change, so do not build a product on it without testing.
- Old versions do not vanish on release day. They stay available for a while, with an announced retirement date.
Two figures appear everywhere: the context window, how much text the model can read at once, and the price per million tokens, charged separately for input and output.
Anthropic: Claude
| Model | Best for | Context |
|---|---|---|
| Claude Fable 5.1 | The most capable: very demanding reasoning and agents that work for hours | 1M |
| Claude Opus 5.5 | Anthropic’s recommended choice for most work; matches Fable 5.1 on most tasks for less than half the price | 1M |
| Claude Sonnet 5.5 | Balance of speed and quality; the model behind Claude.ai | 1M |
| Claude Haiku 4.5 | Fastest and cheapest | 200K |
Opus 5.5 versus Opus 5: Anthropic says it costs 40% less on typical work, writes over 30% faster and improves at coding and computer use. See our Opus 5.5 and Sonnet 5.5 story. Earlier versions such as Fable 5, Opus 5 and Sonnet 5 remain available.
OpenAI: GPT-6
| Model | Best for | Context |
|---|---|---|
| GPT-6 Astra | The most capable, for complex problems where mistakes are costly | 1.05M |
| GPT-6.1 Sol | Coding and multi-tool tasks; close to Astra at a fifth of the price | 1.05M |
| GPT-6 Luna | Very cheap and fast for classifying, extracting or summarising at volume | 1.05M |
All three power ChatGPT and Codex, OpenAI’s coding agent. More in our DevDay 2026 story.
Google: Gemini
| Model | Best for | Status |
|---|---|---|
| Gemini 4 Argon | New frontier generation; up to 1 million output tokens | Rolling out |
| Gemini 3.1 Pro | The large model of the previous generation | Preview |
| Gemini 3.8 Flash | Fast and cheap, with a promotional price until the end of 2026 | Available |
Details in our Gemini 4 Argon story.
xAI: Grok
Grok 4.7 is the flagship, with a 500K-token context and particularly cheap output. Grok 4.3 is cheaper and reaches 1 million tokens, and Grok Build is the coding line. xAI is part of SpaceX, as is Cursor, so Grok is built into that editor. See Grok 4.7.
Meta: Muse Spark
Meta, known for its open Llama models, now leads with Muse Spark, paid through its API, with a 1M-token context and tuned for code in version 1.2. It powers two products: the personal agent Muse and the coding agent Muse Code.
DeepSeek and Mistral
- DeepSeek (China) offers V4-Pro for agents and code and Flash for volume, both with a 1M-token context and prices far below Western models. It charges half price outside peak hours. Before sending it sensitive data, check where it is processed.
- Mistral (France) is the European alternative. Its main model is Mistral Medium 3.5, multimodal and aimed at agents and code, with weights published under its own licence. Its assistant is called Mistral Vibe.
What about Cursor, Claude Code or Codex?
They are not models but tools that use models. Cursor is a code editor where you choose the model (Claude, GPT, Gemini or Grok). Claude Code and Codex are Anthropic’s and OpenAI’s coding agents, running their own models. What differs is how they work with your code, not the “brain” behind them.
What each model costs per million output tokens in the API, cheapest to most expensive (data from our comparison):
- GPT-6 LunaOpenAI0.50 $
- DeepSeek FlashDeepSeek1.2 $
- Grok 4.3xAI2.5 $
- Gemini 3.8 FlashGoogle3.75 $
- DeepSeek V4-ProDeepSeek3.96 $
- Muse Spark 1.2Meta4.25 $
- Claude Haiku 4.5Anthropic5 $
- Grok 4.7xAI6 $
- Claude Sonnet 5.5Anthropic10 $
- GPT-6.1 SolOpenAI10 $
- Gemini 3.1 Pro (preview)Google12 $
- Claude Opus 5.5Anthropic20 $
- Claude Fable 5.1Anthropic50 $
- GPT-6 AstraOpenAI50 $
Which to choose
| If you need… | Start with |
|---|---|
| Coding or long professional work | Claude Opus 5.5, GPT-6.1 Sol or Grok 4.7, and compare |
| The very best, whatever the cost | Claude Fable 5.1 or GPT-6 Astra |
| Everyday chat and writing | The free or entry-level paid plan of ChatGPT, Claude or Gemini |
| High volume at low cost | GPT-6 Luna, Gemini 3.8 Flash, DeepSeek Flash or Claude Haiku 4.5 |
| A European provider | Mistral |
The golden rule: try two or three on your real task before deciding. Company benchmarks are a guide, but what matters is how the model handles what you actually do.
Frequently asked questions
Which AI model is best?
It depends on the task and budget. For coding and long professional work, the top models (Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7) perform similarly, so test them on your own case. For simple, repetitive tasks a small model is far cheaper and good enough.
Why are there so many versions of the same model?
Companies ship new versions every few months. The number goes up with each improvement (Opus 5 → Opus 5.5), and older ones stay available for a while so software that depends on them does not break overnight.
