From a neuron on paper to today’s frontier
Reviewed · 95 entries · 12 frontier labs
Eighty-three years of artificial intelligence on one line: the ideas that started it, the winters that nearly ended it and the labs racing at the frontier today. Every date links to the paper or announcement it comes from.
The frontier on
The most capable model each lab has announced, as the lab itself describes it. Some still reach only a few organizations. 7 of these 12 were announced in the last 60 days.
- Mistral AIMistral Large 4
- GoogleGemini 4 Argon
- xAIGrok 4.7
- Nex AGINex-N2.5
- OpenAIGPT-6 Astra
- AnthropicClaude Fable 5.1
- Z.aiGLM-5.3
- Alibaba QwenQwen3.8
- Moonshot AIKimi K3
- MiniMaxMiniMax M3
- DeepSeekDeepSeek V4
- MetaMuse Spark
Eighty-three years on one ruler
Each mark is one entry of this page, at its real date; the shaded bands are the eras below. 64 of the 95 entries come after ChatGPT.
Since ChatGPT, zoomed in: 65 entries in 1,412 days.
1943–1955
The question
Before computers could do much, a few people asked whether a machine could think, and wrote down how a neuron might compute.
-
A neuron on paper
Warren McCulloch and Walter Pitts describe neurons as logic gates: the first mathematical model of a neural network.
-
Can machines think?
Alan Turing publishes “Computing Machinery and Intelligence” and proposes the imitation game, later called the Turing test.
1956–1972
A field with a name
A workshop names the field. The first programs learn, play and talk, and promise more than the hardware of the time can give.
-
The field gets its name
A summer workshop at Dartmouth College, proposed in 1955 by McCarthy, Minsky, Rochester and Shannon, gives the field its name: artificial intelligence.
-
The perceptron learns
Frank Rosenblatt describes the perceptron, a machine that learns to classify from examples by adjusting its own weights.
-
ELIZA talks back
Joseph Weizenbaum’s ELIZA imitates a therapist with simple pattern matching, and people confide in it anyway.
1973–2005
Winters
Funding dries up, twice. The ideas behind today’s models — backpropagation, convolutions, memory cells — are worked out anyway.
-
The first winter
The Lighthill report tells the British government that AI has not delivered on its promises. Funding is cut and the first AI winter begins.
-
Backpropagation
Rumelhart, Hinton and Williams show in Nature how networks with hidden layers learn by sending the error backwards through them.
-
Convolutions read zip codes
Yann LeCun’s team at Bell Labs trains a convolutional network to read handwritten zip codes, an idea that returns in 2012.
-
Deep Blue beats Kasparov
IBM’s chess computer wins a six-game match against the world champion, Garry Kasparov.
-
Long short-term memory
Hochreiter and Schmidhuber give recurrent networks a memory cell that keeps information across long sequences.
2006–2016
Learning from data
Large datasets and graphics cards make deep networks practical. Machines beat people at Jeopardy!, then at Go.
-
Deep networks, trainable again
Hinton, Osindero and Teh train deep belief networks one layer at a time, and networks with many layers become practical to train.
-
ImageNet
Fei-Fei Li’s team publishes ImageNet, millions of labelled images that become the benchmark of the next decade.
-
Watson wins Jeopardy!
IBM Watson beats the show’s two greatest champions in a televised match, answering questions asked in plain English.
-
An assistant in the pocket
Apple announces Siri with the iPhone 4S and puts a voice assistant in millions of pockets.
-
AlexNet
Krizhevsky, Sutskever and Hinton train a convolutional network on two graphics cards and win the ImageNet challenge by a wide margin.
-
Learning to play Atari
DeepMind’s DQN learns Atari games from the pixels on the screen and the score alone.
-
Generative adversarial networks
Ian Goodfellow and colleagues pit two networks against each other, one generating and one judging, to produce new images.
-
Attention, for translation
Bahdanau, Cho and Bengio let a translation model look back at the words that matter: the attention mechanism.
-
OpenAI is founded
A non-profit research lab is announced in San Francisco, with the goal of advancing digital intelligence in the way most likely to benefit humanity as a whole.
-
AlphaGo beats Lee Sedol
DeepMind’s AlphaGo wins four of five games against one of the best Go players in the world. Move 37 of the second game becomes famous.
2017–2022
Attention
One architecture, the Transformer, scales from translation to code, and several of the labs that lead today are founded.
-
Attention is all you need
Eight researchers, most of them at Google, propose the Transformer, built on attention alone. Almost every model on this page descends from it.
-
GPT
OpenAI pre-trains a Transformer on unlabelled books, then fine-tunes it for each task: the first generative pre-trained transformer.
-
BERT
Google’s BERT reads text in both directions at once and sets new records on eleven language understanding tasks.
-
GPT-2, released in stages
OpenAI holds back the full 1.5-billion-parameter model over concerns about misuse, and publishes it in stages.
-
GPT-3 learns from a few examples
With 175 billion parameters, GPT-3 performs new tasks from a handful of examples in the prompt, without being retrained.
-
AlphaFold and protein folding
At the CASP14 assessment, DeepMind’s AlphaFold 2 predicts protein structures with accuracy close to experiment, a challenge open for 50 years.
-
DALL·E and CLIP
OpenAI shows a model that draws images from a text description, and CLIP, which matches images to the words that describe them.
-
Anthropic appears
Led by Dario and Daniela Amodei, Anthropic announces a $124 million round to build more reliable, interpretable and steerable AI systems.
-
An AI pair programmer
GitHub Copilot, built on an OpenAI model, starts suggesting whole lines and functions inside the editor.
-
Stable Diffusion, open to all
Stability AI and its partners release an image model whose weights anyone can download and run on a consumer graphics card.
2022–2023
The conversation
ChatGPT puts a language model in front of the public. Within a year, every large lab has one of its own.
-
ChatGPT
OpenAI opens a free research preview of a model that answers in conversation. A language model reaches the general public.
-
LLaMA
Meta releases LLaMA to researchers, a family from 7 to 65 billion parameters. The era of open-weight models starts here.
-
GPT-4
OpenAI’s GPT-4 accepts images as well as text, and scores around the top 10% of test takers on a simulated bar exam.
-
Claude
The same day, Anthropic opens Claude to businesses, through a chat interface and an API.
-
Google DeepMind
Google merges Brain and DeepMind into a single unit, Google DeepMind, led by Demis Hassabis.
-
Llama 2, free for commercial use
Meta, with Microsoft as its partner, releases Llama 2 under a license that allows research and commercial use.
-
Mistral 7B
Paris-based Mistral AI releases its first model under the Apache 2.0 license, and it beats Llama 2 13B on every benchmark the lab reports.
-
The Bletchley Declaration
At the first AI Safety Summit, 28 countries and the European Union sign a declaration on the risks of frontier AI.
-
Grok
xAI, the lab Elon Musk founded that year, announces Grok, a chat model with real-time knowledge of the world through X.
-
Gemini
Google DeepMind introduces Gemini, multimodal from the ground up, in Ultra, Pro and Nano sizes.
2024–2025
Reasoning and agents
Models learn to think before they answer and to use computers and tools, while open models from China and Europe close the gap.
-
Claude 3
Anthropic releases Haiku, Sonnet and Opus; Claude 3 Opus outperforms its peers on most of the common benchmarks of the time.
-
GPT-4o
OpenAI’s “omni” model handles voice, image and text in a single network, answering speech in as little as 232 milliseconds.
-
Claude 3.5 Sonnet and Artifacts
A mid-sized model outperforms Claude 3 Opus, and Artifacts let people work on code and documents beside the chat.
-
Llama 3.1 405B
Meta publishes an open-weight model with 405 billion parameters that competes with the closed frontier models of the time.
-
The EU AI Act enters into force
The first comprehensive law on AI, Regulation (EU) 2024/1689, starts to apply in phases, according to the level of risk.
-
o1 thinks before it answers
OpenAI’s o1-preview spends time reasoning in a hidden chain of thought, and does far better on hard math and science problems.
-
A Nobel Prize for neural networks
John Hopfield and Geoffrey Hinton receive the Nobel Prize in Physics for the discoveries that enable machine learning with artificial neural networks.
-
…and one for AlphaFold
Demis Hassabis and John Jumper share the Nobel Prize in Chemistry with David Baker, for predicting and designing protein structures.
-
Claude uses a computer
An upgraded Claude 3.5 Sonnet learns to look at a screen, move the cursor, click and type, in a public beta.
-
Model Context Protocol
Anthropic open-sources MCP, a standard way to connect AI assistants to the tools and data where work happens.
-
Gemini 2.0, for the agentic era
Google’s new generation starts with 2.0 Flash, with native image and audio output and native tool use.
-
DeepSeek-V3
DeepSeek releases an open mixture-of-experts model with 671 billion parameters, 37 billion of them active for each token.
-
DeepSeek-R1
An open reasoning model under the MIT license, with performance on par with OpenAI’s o1 in math, code and reasoning.
-
Claude 3.7 Sonnet and Claude Code
The first hybrid reasoning model arrives with Claude Code, an agent that works on a codebase from the terminal.
-
Gemini 2.5, a thinking model
Google’s Gemini 2.5 Pro reasons before it answers and debuts at the top of the LMArena leaderboard.
-
Llama 4
Meta releases Llama 4 Scout and Maverick, its first natively multimodal open models built on a mixture of experts.
-
o3 and o4-mini
OpenAI’s reasoning models learn to use every tool in ChatGPT, from web search to Python, inside their chain of thought.
-
Qwen3
Alibaba’s Qwen team opens a whole family, from 0.6 to 235 billion parameters, with thinking and non-thinking modes.
-
Claude 4
Claude Opus 4 and Sonnet 4 keep working on long tasks for hours, and Claude Code becomes generally available.
-
Grok 4
xAI scales reinforcement learning to the size of pre-training on its Colossus cluster to train Grok 4.
-
Kimi K2
Moonshot AI opens the weights of a mixture-of-experts model with a trillion parameters, built for agentic work.
-
GLM-4.5
Zhipu AI, now Z.ai, releases open models that bring reasoning, coding and agentic abilities together.
-
gpt-oss, OpenAI’s open weights
For the first time since GPT-2, OpenAI publishes open-weight language models, with 120 and 20 billion parameters.
-
GPT-5
A unified system that decides when to answer quickly and when to think longer, and becomes the default model in ChatGPT.
-
Claude Sonnet 4.5
Anthropic’s model for coding and agents keeps its focus for more than 30 hours on complex, multi-step tasks.
-
Gemini 3
Google calls it its most intelligent model, and puts it in Search and the Gemini app on the first day.
-
Claude Opus 4.5
Anthropic’s model for coding, agents and computer use arrives at a third of the price of the previous Opus.
-
Mistral 3
Mistral Large 3, a mixture of experts with 675 billion parameters, arrives with the small Ministral 3 models, all under Apache 2.0.
2026
The frontier, now
Releases arrive weeks apart, and the most capable models reach restricted groups before everyone else.
-
Qwen3.5
Alibaba opens a multimodal model with 397 billion parameters, 17 billion of them active for each token, built for agents.
-
Gemini 3.1 Pro
Google’s smarter model for complex tasks reaches the Gemini app, developers and enterprises in preview.
-
Claude Mythos, held back
Anthropic does not release Claude Mythos Preview: it finds software vulnerabilities better than all but the most skilled humans. Through Project Glasswing, only partners that secure critical software get access.
-
Muse Spark
The first model from Meta Superintelligence Labs, after nine months spent rebuilding Meta’s AI stack from the ground up.
-
GPT-5.5
OpenAI’s new flagship reaches ChatGPT; GPT-5.5 and GPT-5.5 Pro arrive in the API the next day.
-
DeepSeek V4
Open weights again: V4-Pro has 1.6 trillion parameters, 49 billion active, and both versions read a 1-million-token context.
-
MiniMax M3
A natively multimodal model with about 428 billion parameters, 23 billion active, and a 1-million-token context.
-
Claude Fable 5 and Mythos 5
Anthropic opens a Mythos-class model to the public as Fable 5, and requests in high-risk areas get answered by Claude Opus 4.8 instead. Mythos 5 stays with Project Glasswing partners.
-
Fable 5 is back
On June 12, U.S. export controls on Fable 5 and Mythos 5 led Anthropic to suspend both for every user. The controls were lifted on June 30, and Fable 5 returns worldwide.
-
GPT-5.6: Sol, Terra and Luna
After a limited preview in June, OpenAI splits a generation into three tiers, from the flagship Sol to the fast, low-cost Luna.
-
Kimi K3
Moonshot AI announces a model with 2.8 trillion parameters, native vision and a 1-million-token context; the full weights follow on July 27.
-
Claude Opus 5
Close to the frontier intelligence of Claude Fable 5 at half the price, and the new default model on Claude Max.
-
Muse Glimmer
Meta returns to open weights with Muse Glimmer 30B, a model for agents under the Apache 2.0 license.
-
Qwen3.8
The Qwen team opens Qwen3.8-2.4T-A95B: 2.4 trillion parameters, 95 billion of them active for each token.
-
GLM-5.3
Z.ai’s flagship for complex coding and long-horizon tasks; its open weights follow later in August.
-
Claude Fable 5.1
Anthropic’s most capable model for coding and knowledge work. Its twin, Mythos 5.1, stays in trusted-access programs.
-
Gemini 3.8 Flash
Google’s fast model arrives with 3.8 Flash Cyber, a security model that only trusted defenders can use.
-
GPT-6 Astra
OpenAI’s most intelligent and aligned model yet, for computer use, coding, cybersecurity and science. A limited set of organizations gets it first, then paid ChatGPT plans and the API.
-
Nex-N2.5
Nex AGI releases the N2.5 family: Mini and Pro bring computer use, and Max is a text model with 1.6 trillion parameters.
-
Grok 4.7
The lab’s most powerful model for coding and knowledge work. Its page on x.ai now signs as SpaceXSI.
-
Claude Opus 5.5
The first model of the 5.5 family performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5.
-
Claude Sonnet 5.5
The second model of the 5.5 family runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work.
-
GPT-6.1 Sol
Announced at DevDay, it nearly matches GPT-6 Astra on agentic coding, computer use and professional work, at a fifth of Astra’s price.
-
Gemini 4 Argon
Google’s new frontier model opens first to trusted cyber defenders; paid API customers and Google AI Ultra subscribers come next.
-
Mistral Large 4
Mistral’s largest model: a trillion parameters, 52 billion active, natively multimodal, in public preview with the weights promised for the end of the month.
-
GPT-6 for everyone
GPT-6 reaches ChatGPT users around the world, together with what OpenAI calls Intelligent UI.
-
Claude Haiku 5.5
Anthropic calls it the cheapest, fastest and most capable small model it has released, and it completes the 5.5 family.
· you are here
The next line has no date yet.
How this page is kept: each date is the day of the announcement or publication, not of availability in every country, and each one links to where it comes from. The entries are reviewed by hand; the review date is at the top.