EN : Choose language

7 model families

Models and their versions

GPT-5, Claude Opus 5.5, Gemini 3.8 Flash… The names keep changing. Here are the successive versions of each major family and what each one actually brought.

Last checked: 14 October 2026

How to read a model name

The number
A whole number (GPT-5, Claude 5, Gemini 3) marks a new generation, often a real leap in capability. A number after the dot (5.1, 5.5) is an improvement within the same generation.
The size
Small names (mini, nano, Flash, Haiku, Luna) are fast, cheap models. Big ones (Pro, Opus, Ultra, Astra) are more capable but slower and more expensive.
Reasoning
“Thinking” and “Deep Think” models, or OpenAI’s o-series, take time to think before answering: better at maths, code and logic, but slower.
Open or closed
An open model (Llama, Mistral, DeepSeek) can be downloaded and run on your own machines. A closed model (GPT, Claude, Gemini) is only available through the provider’s service.

To understand what a model is, read the entry LLM.

GPT

OpenAI · mixed

Latest releaseGPT-6.1 Sol

Names: GPT-number, with variants (mini, nano); in generation 6: Astra (most capable), Sol (coding), Luna (budget). The “o” series (o1, o3) were the reasoning models before they were merged into GPT-5.

Where to find it: ChatGPT, Microsoft Copilot, GitHub Copilot, OpenAI API

  1. GPT-6.1 SolNew

    Announced as close to Astra in programming and computer use for about one fifth of the cost. According to the Wall Street Journal, OpenAI, however, decided not to release GPT-6.1 Astra, judged not safe enough.

  2. GPT-6 Sol & Luna

    Round out the range: Sol targets coding at a moderate price, Luna high volumes at very low cost.

  3. GPT-6 Astraflagship

    A new generation presented as a “generational leap” in coding, professional work, cybersecurity and science. The most capable, and the most expensive.

  4. GPT-5.4

    The first mainline model able to use a computer (click, type, browse). Mini and nano versions; mini becomes available to free accounts.

Show the 9 earlier versions

  1. GPT-5.2

    Three modes to choose from: Instant (fast), Thinking (reasoning) and Pro (deep reasoning).

  2. GPT-5flagship

    Merges the fast model and the reasoning model: the system decides on its own when to think. Fewer hallucinations announced.

  3. gpt-ossopen

    OpenAI’s first open models since GPT-2: downloadable and usable on your own machines.

  4. o3 & o4-minireasoning

    Reasoning can use tools (web search, running code, image analysis) while thinking.

  5. o1reasoning

    The first reasoning model: it “thinks” before answering. Big gains in maths and code, at the cost of slower answers.

  6. GPT-4o

    Natively multimodal: understands and produces text, images and voice, with near-instant voice conversation. Becomes ChatGPT’s free model.

  7. GPT-4 Turbo

    Working memory (context) extended to around 300 pages, more recent knowledge, lower price.

  8. GPT-4flagship

    A leap in reasoning and reliability; OpenAI’s first model able to read images.

  9. GPT-3.5

    The model behind the first ChatGPT: fluent conversation, but frequent mistakes and limited knowledge.

Claude

Anthropic · closed

Latest releaseClaude Haiku 5.5

Names: Several sizes per generation, from lightest to most powerful: Haiku, Sonnet, Opus, then since 2026 Fable (public) and Mythos (restricted access).

Where to find it: claude.ai, Claude Code, GitHub Copilot, Amazon Bedrock, Google Vertex AI

  1. Claude Haiku 5.5New

    The smallest and cheapest of the 5.5 family: $0.10 per million input tokens for short requests, about 75% cheaper to run than Haiku 4.5 according to Anthropic; built for simple, repeated tasks.

  2. Claude Sonnet 5.5

    Over 30% faster than Sonnet 5 and up to 30% cheaper per task, according to Anthropic; good at everyday tasks, fixing bugs and producing documents, slides and spreadsheets.

  3. Claude Opus 5.5

    Matches Fable 5.1 on most tasks, at a running cost about 40% lower than Opus 5. Built for coding and long tasks.

  4. Claude Fable 5.1

    An update of the flagship model, released together with Mythos 5.1.

Show the 9 earlier versions

  1. Claude Sonnet 5 & Opus 5

    Opus 5 approaches Fable 5’s capabilities for about half the price, with a one-million-token context.

  2. Claude Fable 5 & Mythos 5flagship

    A new tier above Opus. Fable is open to the public with stronger safeguards in cybersecurity and biology; Mythos, less restricted, is initially reserved for Project Glasswing partners.

  3. Claude Sonnet 4.5 & Opus 4.5

    Marked progress in coding and agents; Haiku 4.5 offers performance close to earlier large models at a small price.

  4. Claude 4 (Sonnet 4, Opus 4)flagship

    Designed for long autonomous tasks and programming, able to work for hours at a time.

  5. Claude 3.7 Sonnetreasoning

    The first “hybrid” model: quick answer or extended thinking, as you choose. Claude Code launches.

  6. Claude 3.5 Sonnet

    The mid-size model beats the previous Opus for much less, with a clear leap in coding. Artifacts arrive; in October, an experimental ability to use a computer.

  7. Claude 3 (Haiku, Sonnet, Opus)flagship

    Introduces the three sizes and image understanding. Opus reaches the level of the best models of the time.

  8. Claude 2

    Opened to the public; reads very long documents (around 75,000 words).

  9. Claude

    Anthropic’s first assistant, built with a method of aligning the model to written principles (“Constitutional AI”).

Gemini

Google · closed

Latest releaseGemini 4 Argon

Names: Gemini-number then the size: Flash-Lite (lightest), Flash (fast), Pro (most capable). “Deep Think” is the deep reasoning mode, “Live” real-time voice conversation.

Where to find it: Gemini app, Google Search, Gmail and Docs, Android, Google API

  1. Gemini 4 ArgonNewflagship

    A new flagship model, first reserved for cyber-defence teams, then opened on 13 October to developers and Pro and Ultra subscribers, at an introductory price.

  2. Gemini 3.8 Flash & Live

    Flash improves at programming, long-running agents and multi-step reasoning; Live is dedicated to low-latency voice conversation.

  3. Gemini 3.5 Flash

    Unveiled at Google I/O in May; AI computer use was built in from June. Flash-Lite follows in July.

  4. Gemini 3.1 Pro

    An update of the Pro model.

Show the 5 earlier versions

  1. Gemini 3 (Pro, Deep Think)flagship

    A new generation focused on reasoning and agentic coding; Flash follows in December.

  2. Gemini 2.5reasoning

    Built-in reasoning (“thinking”); tops many leaderboards at launch.

  3. Gemini 2.0

    Designed for agents; natively generates images and voice. Flash becomes the default model.

  4. Gemini 1.5

    A one-million-token context: can read a whole book or an hour of video at once.

  5. Gemini 1.0flagship

    Designed from the start for text, images, audio and video. Replaces PaLM behind Bard.

Llama, then Muse

Meta · mixed

Latest releaseMuse Wonder

Names: Llama-number with the size in billions of parameters (8B, 70B, 405B). In 2026, Meta launches a closed range, Muse.

Where to find it: Meta AI (WhatsApp, Instagram, Facebook), Hugging Face, Ollama

  1. Muse WonderNewflagship

    A model for agent tasks (Muse Wonder and Muse Wonder Full), available through Meta’s API and the Muse Code tool.

  2. Muse Spark 1.3

    An update focused on programming and long-running agents.

  3. Muse Sparkflagship

    Meta Superintelligence Labs replaces Llama with a closed range: for now, the end of Meta’s large open models.

  4. Llama 4 (Scout, Maverick)open

    “Mixture of experts” architecture, multimodal, very long context. A mixed reception against the competition.

Show the 4 earlier versions

  1. Llama 3.2open

    Adds vision and small models able to run on a phone.

  2. Llama 3 / 3.1open

    Much stronger; in July, the 405B version rivals closed models, with a long context.

  3. Llama 2open

    A licence allowing commercial use: companies can run it on their own servers.

  4. LLaMAopen

    Shared with researchers; its weights leaked and kick-started the open-model ecosystem.

Mistral

Mistral AI · mixed

Latest releaseMistral Large 4

Names: Small, Medium, Large by size; Codestral for code, Magistral for reasoning. Many models are released with open weights.

Where to find it: Le Chat, Mistral API, Hugging Face, Azure, on-premises deployments

  1. Mistral Large 4Newflagship

    Released as a preview: about 1 trillion parameters (52 billion active), natively multimodal, trained in Mistral’s European data centres. Open weights announced for the end of October.

  2. Mistral Small 4

    A new generation of the small model.

  3. Mistral Large 3open

    A large open model (675 billion parameters, 41 billion active), reads images, 256,000-token context.

  4. Magistralreasoning

    Mistral’s first reasoning model, able to reason in French and other languages.

Show the 5 earlier versions

  1. Mistral Small 3open

    A capable small model under the Apache 2.0 licence, free to use.

  2. Mistral Large 2

    Big progress in code, maths and languages.

  3. Mistral Largeflagship

    The first large commercial model; Le Chat launches and a partnership with Microsoft is signed.

  4. Mixtral 8x7Bopen

    Popularises the “mixture of experts” architecture: only part of the model works on each word, making it fast and economical.

  5. Mistral 7Bopen

    A small open-licence model that beats much bigger ones: the French start-up’s breakthrough.

DeepSeek

DeepSeek · open

Latest releaseDeepSeek V4.1 Flash

Names: V-number for general models, R for reasoning (merged since V3.1); Flash (light) and Pro (large) since V4.

Where to find it: DeepSeek app, API, Hugging Face, many hosting providers

  1. DeepSeek V4.1 FlashNew

    Replaces V4 Flash in the API, with a one-million-token context.

  2. DeepSeek V4 (Flash, Pro)flagship

    A new generation under the MIT licence: Flash is light and cheap, Pro a very large model (1.6 trillion parameters).

  3. DeepSeek V3.1

    A single “hybrid” model that can answer quickly or reason.

  4. DeepSeek R1reasoning

    An open reasoning model (MIT licence) close to o1: it shakes the stock market.

Show the 2 earlier versions

  1. DeepSeek V3open

    A large open model on par with the best, reportedly trained for a fraction of the usual cost.

  2. DeepSeek V2open

    High performance at a very low price, triggering a price war in China.

Grok

xAI · closed

Latest releaseGrok 4.7

Names: Grok-number; some older versions were released with open weights.

Where to find it: X (formerly Twitter), Grok app, xAI API, GitHub Copilot

  1. Grok 4.7New

    Context of up to 500,000 tokens, low price; available in Cursor and GitHub Copilot among others.

  2. Grok 4flagship

    Aims for the top of reasoning and tool-use leaderboards.

  3. Grok 3reasoning

    Trained on a very large supercomputer; adds a reasoning mode.

  4. Grok 2

    Major progress and image generation.

Show the 1 earlier versions

  1. Grok 1

    xAI’s first model, built into X, with a deliberately irreverent tone.

New versions are tracked every week in the news.

See the news

Updated on 14 October 2026. Recent versions rely partly on secondary sources and are being reviewed.