Skip to main content

From 1950 to 2026

The History of AI

From Turing’s question “Can machines think?” to agents that program on their own for hours. Six eras, 139 sourced milestones and a clearly marked outlook.

Editorial review:

The six eras at a glance

1950-2011

The foundations

Turing, Dartmouth and the first learning networks. Decades of research before computing power and data were enough.

Key moments

University of Manchester

Alan Turing asks: Can machines think?

Alan Turing describes the imitation game, later known as the Turing test.

Read more

The paper shifts the debate from defining thought to observable behavior.

Text & LanguageSource
Dartmouth College

The Dartmouth workshop coins the term AI

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon bring researchers together at Dartmouth.

Read more

Their earlier project proposal uses the term 'Artificial Intelligence' and gives the field its name.

Text & LanguageSource

More milestones4, in date order

The perceptron: learning with artificial neuronsCornell Aeronautical Laboratory · 1958

Frank Rosenblatt publishes the perceptron, an early trainable neural network. It links mathematical neuron models with a trainable classifier.

ImageResearchSource
Backpropagation makes multilayer networks trainableUniversity of California San Diego · 1986

Rumelhart, Hinton, and Williams show how errors can be propagated through multiple network layers. The algorithm becomes a practical core of deep learning.

ImageResearchSource
LSTM: memory for long sequencesTechnical University of Munich · 1997

Sepp Hochreiter and Jürgen Schmidhuber introduce Long Short-Term Memory. The architecture helps networks retain relevant information over longer sequences and shapes language and audio AI before Transformers.

Text & LanguageResearchSource
ImageNet creates a shared test for image recognitionPrinceton University · 2009

ImageNet publishes a large, hierarchically labelled image dataset. Its later challenge makes progress in image classification measurable and sets the stage for AlexNet.

ImageResearchSource

2012-2016

Deep learning takes over

AlexNet starts the rise of deep neural networks. In 2016, AlphaGo shows what they can do.

Key moments

University of Toronto

AlexNet wins ImageNet

The AlexNet neural network wins the ImageNet competition by a wide margin, slashing the error rate.

Read more

The victory ignites the deep learning revolution.

ImageSource
Université de Montréal

GANs: networks that learn against each other

Ian Goodfellow introduces generative adversarial networks.

Read more

Two networks train against each other to produce strikingly realistic images, the basis for many later image generators.

ImageSource
OpenAI

OpenAI is founded

Elon Musk, Sam Altman, and others found OpenAI as a non-profit research lab with the stated goal of developing AI to benefit all of humanity.

Read more
MultimodalSource
DeepMind

AlphaGo beats Lee Sedol

DeepMind's AlphaGo beats world-class Go player Lee Sedol 4-1. The game long thought unwinnable for machines falls to an AI years earlier than expected.

Read more
Agents & ToolsSource

More milestones7, in date order

word2vec: words become vectorsGoogle · Jan 2013

A Google team led by Tomáš Mikolov introduces word2vec. Words become vectors whose distances capture meaning, a foundation of modern language AI.

Text & LanguageResearchSource
Google acquires DeepMindGoogle · Jan 2014

Google acquires the London AI lab DeepMind. The deal shows that foundational AI research has become strategic for the major platform companies.

Agents & ToolsMilestone
Sequence-to-sequence learningGoogle · Sep 2014

Google researchers demonstrate sequence-to-sequence learning with neural networks. The encoder-decoder principle becomes the foundation for machine translation and later the Transformers.

Text & LanguageResearchSource
Deep Q-Network masters AtariDeepMind · Feb 2015

DeepMind's Deep Q-Network learns 49 Atari games from pixels and score alone, many at human level. For the first time a network masters such tasks through pure trial and error.

Agents & ToolsResearchSource
TensorFlow goes open sourceGoogle · Nov 2015

Google releases TensorFlow as open source. An industrial-grade framework makes training neural networks accessible to everyone.

Agents & ToolsToolOpen WeightsSource
ResNet: networks go deepMicrosoft · Dec 2015

Microsoft Research introduces residual networks. "Skip connections" enable extremely deep networks and win the 2015 ImageNet competition.

ImageResearchSource
WaveNet: natural-sounding voicesDeepMind · Sep 2016

DeepMind's WaveNet generates speech directly as raw audio, sounding far more natural than earlier synthesis. The technique later ships in Google products.

Audio & MusicResearchSource

2017-2021

Transformers and large language models

A new architecture makes language models scalable. GPT-3 and Copilot show where this leads.

Key moments

Google

Transformer: "Attention Is All You Need"

A Google team introduces the Transformer architecture, built purely on attention.

Read more

It becomes the foundation for GPT, BERT, and every modern language model.

Text & LanguageSource
OpenAI

GPT-3: 175 billion parameters

OpenAI releases GPT-3, demonstrating emergent abilities like coding, translation, and reasoning without task-specific training.

Read more
Text & LanguageProprietarySource
GitHub

GitHub Copilot: AI writes alongside you

GitHub and OpenAI launch Copilot as a technical preview.

Read more

It suggests code directly in the editor and brings generative AI into day-to-day software development.

Agents & ToolsProprietarySource

More milestones5, in date order

GPT points the way to pretrained language modelsOpenAI · Jun 2018

OpenAI publishes the first Generative Pre-trained Transformer model. The approach combines unsupervised pretraining with later adaptation to specific language tasks.

Text & LanguageResearchSource
BERT improves language understandingGoogle · Oct 2018

Google introduces BERT. The bidirectionally trained Transformer model shapes search systems and many language-understanding tasks.

Text & LanguageResearchSource
GPT-2: "too dangerous to release"OpenAI · Feb 2019

OpenAI shows GPT-2 with 1.5 billion parameters but initially withholds the full weights. For the first time, machine-written text reads consistently fluent.

Text & LanguageModelOpen WeightsSource
DALL-E: text becomes imageOpenAI · Jan 2021

OpenAI introduces DALL-E, generating images from text descriptions. Generative AI becomes tangible for a broad audience for the first time.

ImageResearchSource
Google introduces the dialogue model LaMDAGoogle · May 2021

Google presents LaMDA as a language model for open-ended conversation. The later sentience debate shows how strongly fluent dialogue can shape public perceptions of language models.

Text & LanguageModelProprietarySource

2022-2023

Generative AI for everyone

ChatGPT, Stable Diffusion and GPT-4 bring AI into everyday life. Open models follow.

Key moments

Stability AI

Stable Diffusion: open weights for images

Stability AI releases Stable Diffusion with open weights.

Read more

High-quality image generation suddenly runs free on your own GPU.

ImageOpen WeightsSource
OpenAI

ChatGPT: AI for the masses

OpenAI releases ChatGPT as a conversational interface for a language model.

Read more

Its unusually rapid adoption makes generative AI visible to a broad audience.

Text & LanguageProprietarySource
OpenAI

GPT-4: human-level performance on exams

OpenAI releases GPT-4, which understands images and tops bar and medical exams.

Read more

It becomes the benchmark for two years.

MultimodalProprietarySource
Meta

Llama 2: open source goes commercial

Meta releases Llama 2 for free commercial use. A turning point that sparks an entire industry of open-weight models.

Read more
Text & LanguageOpen WeightsSource
Mistral AI

Mistral 7B: small, fast, surprisingly strong

Young Paris startup Mistral shows with 7 billion parameters that small open models can beat much larger ones.

Read more
Text & LanguageOpen Weights

More milestones25, in date order

20226 entries

DALL-E 2: photorealistic image synthesisOpenAI · Apr 2022

DALL-E 2 raises image quality sharply and popularizes diffusion models. The text-to-image race becomes visible to everyone.

ImageModelProprietarySource
Google PaLM: 540 billion parametersGoogle · Apr 2022

With PaLM, Google shows how far pure scaling reaches, reporting strong results on reasoning and multilingual tasks.

Text & LanguageModelProprietarySource
Midjourney opens its public betaMidjourney · Jul 2022

Midjourney makes text-to-image publicly accessible through a Discord bot. Its distinct, stylized look quickly becomes a signature.

ImageProductProprietary
Show 3 more from 2022
Meta Make-A-Video: first text-to-video stepsMeta · Sep 2022

Meta shows Make-A-Video, generating short clips from text prompts. The results are rough but mark the start of video AI.

VideoResearchSource
OpenAI Whisper: robust speech recognitionOpenAI · Sep 2022

OpenAI open-sources Whisper. Accurate transcription across many languages becomes a freely available foundation for audio tools.

Audio & MusicModelOpen WeightsSource
Perplexity: an answer engine, not a link listPerplexity · Dec 2022

Perplexity launches as an AI search engine that answers questions directly with cited sources. The format challenges classic web search.

Text & LanguageProductProprietary

202319 entries

ElevenLabs: uncannily real AI voicesElevenLabs · Jan 2023

ElevenLabs opens its beta for speech synthesis and voice cloning. The naturalness of the voices sets a new bar for audio AI.

Audio & MusicProductProprietary
Microsoft Bing Chat: GPT-4 in searchMicrosoft · Feb 2023

Microsoft builds a GPT-4 model into Bing. The chatbot draws attention with its quirky replies and opens the search-engine race.

Text & LanguageProductProprietary
Meta LLaMA: research goes (almost) freeMeta · Feb 2023

Meta releases LLaMA to researchers; soon after, the weights circulate freely. The open-source community builds its own language models overnight.

Text & LanguageModelOpen Weights
Show 16 more from 2023
Adobe Firefly: image AI for professionalsAdobe · Mar 2023

Adobe brings Firefly straight into Photoshop and stresses commercially safe training data. Generative AI enters the professional workflow.

ImageProductProprietary
Google Bard: the answer to ChatGPTGoogle · Mar 2023

Google launches Bard as its own chatbot. The bumpy debut shows how much pressure ChatGPT put on the market leader.

Text & LanguageProductProprietary
Anthropic Claude: the safety-first alternativeAnthropic · Mar 2023

Anthropic introduces the first Claude, positioning itself with "Constitutional AI" as a safety-focused alternative to OpenAI.

Text & LanguageProductProprietary
Midjourney v5: photorealism becomes normalMidjourney · Mar 2023

With v5, Midjourney produces images barely distinguishable from photos. The first AI images go viral and fool millions.

ImageModelProprietary
Runway Gen-2: text becomes videoRunway · Jun 2023

Runway makes text-to-video publicly usable. The clips are short and flickery, but for the first time anyone can generate moving images.

VideoProductProprietary
Claude 2: 100,000-token contextAnthropic · Jul 2023

Anthropic releases Claude 2 with a then-huge 100,000-token context window. Entire books fit into a single prompt.

Text & LanguageModelProprietary
Stable Diffusion XL: sharper open imagesStability AI · Jul 2023

Stability AI follows up with SDXL, markedly improving composition and detail in open image models.

ImageModelOpen Weights
Code Llama: an open model for codingMeta · Aug 2023

Meta specializes Llama 2 for code. Open-source development gains a freely usable programming assistant.

Agents & ToolsModelOpen Weights
Falcon 180B: the largest open modelTII · Sep 2023

Abu Dhabi's Technology Innovation Institute releases Falcon 180B, briefly topping the open-model leaderboard.

Text & LanguageModelOpen Weights
DALL-E 3: image AI that understands real promptsOpenAI · Oct 2023

DALL-E 3 moves into ChatGPT and follows complex instructions far more closely. Image generation becomes a conversation.

ImageModelProprietary
GPT-4 Turbo and the first "GPTs"OpenAI · Nov 2023

At its first DevDay, OpenAI shows GPT-4 Turbo with a 128,000-token context and custom configurable GPTs. The agent idea takes shape.

Text & LanguageProductProprietary
xAI Grok: Elon Musk's chatbotxAI · Nov 2023

Elon Musk's xAI introduces Grok, tied closely to X and deliberately cheeky in tone. A few months later the weights are opened.

Text & LanguageModelOpen Weights
Pika 1.0: video AI for everyonePika · Nov 2023

Startup Pika brings accessible video generation and raises substantial venture capital. The video-AI race speeds up.

VideoProductProprietary
Google Gemini: natively multimodalGoogle · Dec 2023

Google introduces Gemini, built from the ground up for text, image, audio, and video. It rivals GPT-4 across many benchmarks.

MultimodalModelProprietary
Midjourney v6: readable text in imagesMidjourney · Dec 2023

With v6, Midjourney reliably renders text inside images for the first time and pushes photographic quality further.

ImageModelProprietary
Mixtral 8x7B: mixture of experts goes mainstreamMistral AI · Dec 2023

Mistral releases Mixtral, showing how a mixture-of-experts model delivers strong performance at lower cost per query.

Text & LanguageModelOpen Weights

The era in data

Context-window records (log)
Best value
OpenAI
Anthropic
Google
Meta
Source: vendor documentation
gradually.ai

A large context window describes the maximum input length, not the quality across the whole context.

2024-2025

Multimodal and reasoning

Models see, hear and think for longer. On GPQA Diamond, the best score rises from 50.4% to 91.7%.

Key moments6 of 19 as cards, the other 13 carry a star in the list

OpenAI

OpenAI Sora: a shock to the video world

OpenAI shows Sora demos with photorealistic clips up to a minute long.

Read more

The quality is years ahead of rivals and sparks a debate.

VideoSource
Anthropic

Claude 3: Anthropic surpasses GPT-4

Anthropic releases Claude 3 (Opus, Sonnet, Haiku), beating GPT-4 on many benchmarks.

Read more

For the first time, OpenAI is no longer unchallenged.

MultimodalProprietary
Meta

Llama 3.1 405B: open at frontier level

Meta releases Llama 3.1 405B, the first open model to keep pace with the best proprietary systems.

Read more
Text & LanguageOpen WeightsSource
Black Forest Labs

FLUX.1: a new open image flagship

Black Forest Labs, founded by Stable Diffusion's inventors, releases FLUX.1 and sets a new standard for open image models.

Read more
ImageOpen Weights
DeepSeek

DeepSeek-V3: frontier performance for $5.6M

DeepSeek trains an open 671-billion model for about $5.6 million, calling into question the entire industry's cost assumptions.

Read more
Text & LanguageOpen WeightsSource
Google

Gemini 2.5 Pro: Google takes the lead

Google releases Gemini 2.5 Pro, topping many leaderboards clearly for the first time.

Read more

The company is back in the race for the best model.

MultimodalProprietary

More milestones41, in date order

202419 entries

Gemini 1.5 Pro: 1 million-token contextGoogle · Feb 2024

Google shows Gemini 1.5 Pro with a one-million-token context window. Whole codebases or hours of video fit into a single prompt.

MultimodalModelProprietary
Cognition Devin: the "first AI software engineer"Cognition · Mar 2024

Cognition introduces Devin, an agent that plans and executes whole coding tasks on its own. The vision of autonomous coding agents becomes tangible.

Agents & ToolsProductProprietary
Suno v3: complete songs at the push of a buttonSuno · Mar 2024

Suno turns a short prompt into finished songs with vocals and instruments. AI music becomes mainstream overnight.

Audio & MusicProductProprietary
Show 16 more from 2024
Llama 3: open source closes the gapMeta · Apr 2024

Meta releases Llama 3 with performance close to proprietary models. Open AI becomes truly competitive.

Text & LanguageModelOpen WeightsSource
Udio: the second major music generatorUdio · Apr 2024

Udio launches with especially clean audio quality and challenges Suno. The AI music market becomes a duel.

Audio & MusicProductProprietary
GPT-4o: real-time voice and multimodalityOpenAI · May 2024

OpenAI ships GPT-4o, processing text, image, and audio in real time. Natural spoken conversation with AI becomes reality.

MultimodalModelProprietarySource
Claude 3.5 Sonnet and ArtifactsAnthropic · Jun 2024

Anthropic releases Claude 3.5 Sonnet, beating Claude 3 Opus at lower cost, and introduces "Artifacts" for interactive output.

Text & LanguageModelProprietary
Kling: China's answer to SoraKuaishou · Jun 2024

Chinese company Kuaishou launches Kling, showing that top-tier video models do not come only from the West.

VideoProductProprietary
Luma Dream Machine: video AI for everyoneLuma AI · Jun 2024

Luma makes high-quality video generation freely accessible. Demand is so high the servers buckle at times.

VideoProductProprietary
Runway Gen-3 Alpha: more cinematic clipsRunway · Jun 2024

Runway answers Sora with Gen-3 Alpha, delivering markedly more coherent, cinematic video.

VideoModelProprietary
Stable Diffusion 3: a new open architectureStability AI · Jun 2024

Stability AI releases Stable Diffusion 3 with better prompt adherence and typography. Licensing questions, however, spark debate.

ImageModelOpen Weights
Grok 2: xAI catches upxAI · Aug 2024

xAI releases Grok 2 with built-in image generation, moving close to the top tier of chatbots.

MultimodalModelProprietary
NotebookLM Audio Overviews: AI podcastsGoogle · Sep 2024

With NotebookLM, Google turns your own documents into surprisingly natural podcast dialogues. The feature becomes a viral hit.

Audio & MusicProductProprietary
OpenAI o1: reasoning as a new paradigmOpenAI · Sep 2024

OpenAI releases o1, which "thinks" before answering. Spending more compute at inference time becomes a new source of capability gains.

Text & LanguageModelProprietary
Claude Computer Use: AI controls the computerAnthropic · Oct 2024

Anthropic gives Claude 3.5 Sonnet the ability to operate mouse and keyboard. The agent can use software like a human.

Agents & ToolsProductProprietary
Gemini 2.0 Flash: fast and agenticGoogle · Dec 2024

Google ships Gemini 2.0 Flash with native image and audio output, focusing on agentic applications.

MultimodalModelProprietary
Llama 3.3 70B: smaller, nearly as strongMeta · Dec 2024

With Llama 3.3, Meta shows a 70-billion model nearly matching the five-times-larger 405B model.

Text & LanguageModelOpen Weights
Sora ships, ten months after the demoOpenAI · Dec 2024

OpenAI finally makes Sora publicly available. By now rivals have caught up, and the February lead has melted away.

VideoProductProprietary
Google Veo 2: physically plausible videoGoogle · Dec 2024

Google shows Veo 2 with a better grasp of physics and camera motion, positioning directly against Sora.

VideoModelProprietary

202522 entries

Cursor becomes the default AI editorCursor · Jan 2025

AI code editor Cursor races to hundreds of millions in revenue. Coding with an AI agent inside the editor becomes normal for many teams.

Agents & ToolsProductProprietary
DeepSeek-R1: the "DeepSeek shock"DeepSeek · Jan 2025

Chinese startup DeepSeek releases an open reasoning model on par with o1. The news sends AI stocks tumbling and reshapes the global landscape.

Text & LanguageModelOpen Weights
OpenAI Operator: AI clicks in the browserOpenAI · Jan 2025

OpenAI launches Operator, an agent that operates websites on its own and handles tasks like bookings. Agentic AI leaves the lab.

Agents & ToolsProductProprietary
Show 19 more from 2025
Claude 3.7 Sonnet: first hybrid reasoning modelAnthropic · Feb 2025

Anthropic combines fast answers and visible thinking in one model. Users control how long Claude "thinks" about a task.

Text & LanguageModelProprietary
Claude Code: agentic coding in the terminalAnthropic · Feb 2025

Anthropic introduces Claude Code, an agent that handles whole tasks across many files in the terminal. It becomes the template for agentic coding.

Agents & ToolsProductProprietary
GPT-4.5: the end of pure scalingOpenAI · Feb 2025

OpenAI ships the large, expensive GPT-4.5. The modest gains show that more parameters alone barely move the needle anymore; focus shifts to reasoning.

Text & LanguageModelProprietary
Grok 3: ten times the training computexAI · Feb 2025

xAI trains Grok 3 on a massive GPU cluster with ten times the compute of Grok 2, closing in on the leaders.

Text & LanguageModelProprietarySource
ChatGPT image generation: the "Ghibli" momentOpenAI · Mar 2025

OpenAI builds native image generation into GPT-4o. Millions turn photos into anime styles, and the servers hit their limits for days.

ImageProductProprietary
Llama 4: Meta's mixture-of-experts generationMeta · Apr 2025

Meta introduces Llama 4 Scout and Maverick as open MoE models, with Scout offering a ten-million-token context window.

MultimodalModelOpen WeightsSource
OpenAI o3 and o4-mini: reasoning with toolsOpenAI · Apr 2025

OpenAI releases o3 and o4-mini, which use tools like web search and code on their own while thinking. Agentic reasoning becomes standard.

Text & LanguageModelProprietarySource
Claude Opus 4 and Sonnet 4: agentic codingAnthropic · May 2025

Anthropic releases Claude Opus 4 and Sonnet 4, able to work on code autonomously for hours. Claude takes the lead in programming.

Text & LanguageModelProprietary
Google Veo 3: video with synchronized audioGoogle · May 2025

Veo 3 generates clips with matching sound and dialogue for the first time. Generated video becomes hard to tell from real footage.

VideoModelProprietary
Grok 4: xAI at the benchmark frontierxAI · Jul 2025

xAI releases Grok 4, reporting top scores on several reasoning benchmarks. The Colossus GPU cluster pays off.

Text & LanguageModelProprietary
GPT-5: one model for everythingOpenAI · Aug 2025

OpenAI ships GPT-5, automatically blending fast answers and deep reasoning based on the task. Manual model switching disappears.

MultimodalModelProprietarySource
Nano Banana: image editing by languageGoogle · Aug 2025

Google's image model nicknamed "Nano Banana" edits photos consistently across multiple steps. Targeted editing instead of regenerating becomes the norm.

ImageModelProprietary
Claude Sonnet 4.5: hours-long agent runsAnthropic · Sep 2025

Anthropic releases Claude Sonnet 4.5, which by its own account can work autonomously on complex tasks for over 30 hours.

Text & LanguageModelProprietary
Sora 2 and its own social appOpenAI · Sep 2025

OpenAI ships Sora 2 with synchronized audio and a dedicated app for AI videos. Cameos let users place themselves into generated clips.

VideoProductProprietary
Claude Opus 4.5: new flagshipAnthropic · Nov 2025

Anthropic releases Claude Opus 4.5, cementing its position on agentic and coding tasks.

Text & LanguageModelProprietary
DeepSeek-V3.2: cheap open frontier performanceDeepSeek · Dec 2025

DeepSeek updates its open MoE model and keeps the gap to proprietary leaders small, at a fraction of the cost.

Text & LanguageModelOpen Weights
Gemini 3 Pro: Google's next big leapGoogle · Dec 2025

Google releases Gemini 3 Pro with markedly improved reasoning and multimodal understanding, setting new highs.

MultimodalModelProprietary
GPT-5.2: faster answers on the GPT-5 baseOpenAI · Dec 2025

OpenAI follows up with GPT-5.2, improving speed and reasoning over the original GPT-5.

MultimodalModelProprietary
Mistral Large 3: Europe's flagship modelMistral AI · Dec 2025

Paris-based Mistral releases a 675-billion MoE model, keeping the European flag flying in the AI race.

Text & LanguageModelProprietary

The era in data

Expert knowledge: GPQA Diamond over time
Best value
OpenAI
Anthropic
Google
DeepSeek
Moonshot AI
Source: vendors and Vals AI (from 2025)
gradually.ai

GPQA Diamond tests graduate-level science knowledge. The values come from different configurations and evaluations, not a normalized ranking.

Programming: SWE-bench Verified over time
Best value
Anthropic
OpenAI
Google
DeepSeek
Source: vendors and Vals AI (from 2025)
gradually.ai

SWE-bench Verified starts in August 2024. Scaffolds and compute budgets differ per result.

2026

The year of agents

Models work on tasks for hours. With 37 entries, 2026 is the densest year of this timeline.

Key moments6 of 23 as cards, the other 17 carry a star in the list

Anthropic

Claude 4.6: 1 million tokens and agent teams

Anthropic releases Opus 4.6 and Sonnet 4.6 with a 1-million-token context.

Read more

Opus 4.6 hits 76% on the MRCR v2 benchmark and coordinates "agent teams".

Text & LanguageProprietarySource
Google

Gemini 3.1 Pro: a double reasoning leap

Google releases Gemini 3.1 Pro with more than double the reasoning performance over Gemini 3 Pro and a record 94.3% on GPQA Diamond.

Read more
MultimodalProprietarySource
DeepSeek

DeepSeek-V4-Pro: 1.6 trillion parameters, open

DeepSeek releases an open MoE model with 1.6 trillion parameters (49 billion active) under an MIT license.

Read more

Open models reach trillion scale.

Text & LanguageOpen WeightsSource
OpenAI

GPT-5.5: agentic workflows over hours

GPT-5.5 plans and uses tools autonomously across long-running tasks, reaches 82.7% on Terminal-Bench 2.0, and consumes far fewer tokens doing so.

Read more
Agents & ToolsProprietarySource
Moonshot AI

Kimi K3: a new frontier multimodal model

Moonshot AI initially releases Kimi K3 as an API model.

Read more

The frontier model combines text with native visual capabilities.

MultimodalProprietarySource
Alibaba

Qwen 3.8: open weights for the frontier

Alibaba publishes the Qwen 3.8 weights. The release marks another open-weights milestone for frontier models.

Read more
Text & LanguageOpen WeightsSource

More milestones31, in date order

January to March2 entries

GPT-5.3-Codex: coding and reasoning unifiedOpenAI · Feb 2026

OpenAI unifies frontier coding and professional reasoning in a single model for the first time. It runs about 25% faster than its predecessor.

Agents & ToolsModelProprietarySource
GPT-5.4: OpenAI keeps the paceOpenAI · Mar 2026

OpenAI ships GPT-5.4, hitting 92.8% on GPQA Diamond with gains in knowledge and multilingual tasks.

MultimodalModelProprietary

April to June11 entries

Claude Opus 4.7: adaptive thinking and task budgetsAnthropic · Apr 2026

Anthropic releases Opus 4.7 with a new tokenizer, adaptive thinking, task budgets for agents, and higher-resolution vision. SWE-bench Verified: 87.6%.

Text & LanguageModelProprietarySource
Kimi K2.6: an open trillion-scale model from ChinaMoonshot AI · Apr 2026

Moonshot AI releases Kimi K2.6 with 1 trillion parameters, native INT4 quantization, and a 262,000-token context under a modified MIT license.

Text & LanguageModelOpen WeightsSource
Qwen 3.6 Max: Alibaba's trillion-scale MoEAlibaba · Apr 2026

Alibaba ships Qwen 3.6 Max-Preview, a sparse MoE model with around a trillion parameters, an integrated thinking mode, and a 262,000-token context.

Text & LanguageModelProprietarySource
Show 8 more from April to June
Claude Opus 4.8: dynamic workflowsAnthropic · May 2026

Anthropic releases Opus 4.8 with hundreds of parallel subagents per session, effort control, and a 3x cheaper fast mode. SWE-bench Verified rises to 88.6%.

Text & LanguageModelProprietarySource
Gemini 3.5 Flash: a fast all-rounderGoogle · May 2026

At Google I/O 2026, Gemini 3.5 Flash arrives with four times faster output and strong scores on agentic benchmarks.

MultimodalModelProprietarySource
Claude Fable 5 and Mythos 5: the Mythos classAnthropic · Jun 2026

Anthropic launches Fable 5 and Mythos 5, a new class above Opus with a 1-million-token context. Days later, a US export control directive suspends access.

Text & LanguageModelProprietarySource
GPT-5.6: Sol, Terra, and LunaOpenAI · Jun 2026

OpenAI previews the GPT-5.6 family. The flagship, Sol, hits 88.8% on Terminal-Bench 2.1 (Ultra mode 91.9%) and launches first as a limited preview for around 20 partners.

Agents & ToolsModelProprietarySource
Kimi K2.7 Code: an open coding modelMoonshot AI · Jun 2026

Moonshot AI releases Kimi K2.7 Code, an open MoE model with 1 trillion parameters (32 billion active) and a 256,000-token context, with thinking always on.

Agents & ToolsModelOpen WeightsSource
MiniMax M3: 1 million tokens from ShanghaiMiniMax · Jun 2026

Shanghai lab MiniMax ships M3, an API model with a 1-million-token context. Another Chinese provider pushes toward the frontier.

Text & LanguageModelProprietarySource
Mistral Medium 3.5: tuned for codingMistral AI · Jun 2026

Mistral AI updates its Medium line with a coding-focused API model. The parameter count stays undisclosed.

Agents & ToolsModelProprietarySource
Claude Sonnet 5: a new level for agentic workAnthropic · Jun 2026

Anthropic releases Claude Sonnet 5 for complex agentic tasks. The new generation extends the line with a frontier model for professional workflows.

Agents & ToolsModelProprietarySource

July to September18 entries

Grok 4.5: a new generation of agentsxAI · Jul 2026

xAI advances the Grok line with a new agentic model. Grok 4.5 extends the family for complex tasks involving tools.

Agents & ToolsModelProprietarySource
Gemini 3.6 Flash: fast multimodalityGoogle · Jul 2026

Google further expands the Flash family. Gemini 3.6 Flash targets fast multimodal applications.

MultimodalModelProprietarySource
Claude Opus 5: next-generation frontier reasoningAnthropic · Jul 2026

Anthropic positions Claude Opus 5 as a new model for demanding reasoning. It leads the Claude 5 generation in the most complex professional workflows.

Agents & ToolsModelProprietarySource
Show 15 more from July to September
Kimi K3: open weights for multimodalityMoonshot AI · Jul 2026

Moonshot AI releases the Kimi K3 weights. This makes a frontier model with native visual capabilities openly available.

MultimodalModelOpen WeightsSource
Grok 4.6: xAI advances its agentsxAI · Aug 2026

xAI releases Grok 4.6 as its next agentic model version. The Grok line continues to advance at a rapid pace.

Agents & ToolsModelProprietarySource
Gemini 3.7 Flash: the next Flash iterationGoogle · Aug 2026

Google updates its fast multimodal Flash line again. Gemini 3.7 Flash continues the dense summer release cadence.

MultimodalModelProprietarySource
GLM-5.3: Z.ai introduces a new agentZ.ai · Aug 2026

Z.ai releases GLM-5.3 as a new API model for agentic tasks. The model extends the GLM 5 line for professional applications.

Agents & ToolsModelProprietarySource
Qwen 3.8 Flash: multimodal API for coding and agentsAlibaba · Aug 2026

Alibaba releases Qwen 3.8 Flash as a multimodal API model with a native 1-million-token context for deep thinking, visual understanding, coding, and agentic workflows.

MultimodalModelProprietarySource
Qwen 3.8 Flash Next: open multimodal weightsAlibaba · Aug 2026

Alibaba releases Qwen 3.8 Flash Next as a multimodal MoE under the Qwen Community License 1.0. The model activates 6 billion parameters per token and natively supports a 262,144-token context.

MultimodalModelOpen WeightsSource
Claude Fable 5.1 and Mythos 5.1: stronger agentsAnthropic · Sep 2026

Anthropic releases Fable 5.1 as its generally available frontier model and Mythos 5.1 as a restricted Project Glasswing variant. Both offer a 1 million token context window, 128,000 output tokens, and a June 2026 knowledge cutoff.

Agents & ToolsModelProprietarySource
Gemini 3.8 Flash: stronger coding and agentsGoogle · Sep 2026

Google releases Gemini 3.8 Flash as a stable multimodal model. It offers a 1,048,576-token context window, up to 65,536 output tokens, and improves coding and agentic tasks according to Google.

MultimodalModelProprietarySource
Gemini 3.8 Flash Cyber: defensive cybersecurityGoogle · Sep 2026

Google introduces Gemini 3.8 Flash Cyber for vulnerability discovery and automated patching. Access remains limited to vetted defenders in the Fairwind program.

Agents & ToolsModelProprietarySource
Muse Spark 1.3: longer agentic tasks with fewer stepsMeta · Sep 2026

Meta releases Muse Spark 1.3 for agentic and coding tasks. Meta reports about 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2.

MultimodalModelProprietarySource
Qwen 3.8 Max 0902: stronger coding and agent teamsAlibaba · Sep 2026

Alibaba releases a new Qwen 3.8 Max snapshot. The model retains a 1 million-token context window while improving coding, long-horizon autonomous development, agent collaboration, and visual analysis.

MultimodalModelProprietarySource
GPT-6 Astra: OpenAI's new flagship modelOpenAI · Sep 2026

OpenAI introduces GPT-6 Astra as the flagship of its GPT-6 family. According to the provider, it offers a 1.05 million-token context window and has been rolling out in phases since September.

Agents & ToolsModelProprietarySource
Claude Opus 5.5: new leading model, lower priceAnthropic · Sep 2026

Anthropic replaces Claude Opus 5 with Opus 5.5 as its new leading model, with a 1 million-token context window, a June 2026 knowledge cutoff, and pricing of $4 input and $20 output per 1 million tokens, about 40% cheaper than its predecessor.

Agents & ToolsModelProprietarySource
GPT-6 Sol and Luna: cheaper tiers below AstraOpenAI · Sep 2026

OpenAI releases GPT-6 Sol and GPT-6 Luna as successors to the GPT-5.6 Sol and Luna tiers, each with a 1,050,000-token context window. Sol costs $2/$10 per 1 million tokens, Luna $0.10/$0.50. GPT-6 Astra remains OpenAI's best model overall.

Agents & ToolsModelProprietarySource
Claude Sonnet 5.5: the faster complement to Opus 5.5Anthropic · Sep 2026

Anthropic releases Claude Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5, with a 1 million-token context window, a June 2026 knowledge cutoff, and unchanged pricing of $2 input and $10 output per 1 million tokens.

Agents & ToolsModelProprietarySource

The era in data

Price and performance of current modelsOverall index (higher is better) against price per 1M tokens (further left is cheaper)

most attractive area: stronger and cheaper than half of the models shownbest intelligence for the price

100120140$0.20$1$5$20GPT-6 Sol: 108.4, $4GPT-5.6 Sol: 109.3, $8Muse Ember 1.0: 110.5, $0.20Kimi K3.1 Pro: 112.6, $4Claude Fable 5: 112.8, $20Muse Wonder: 113.7, $8Claude Opus 5: 117.3, $10GPT-6.1 Sol: 117.5, $4GPT-6 Astra: 119.6, $20Gemini 4 Argon: 121.2, $4Claude Sonnet 5.5: 121.9, $4Claude Fable 5.1: 124.1, $20Claude Opus 5.5: 125.7, $8Gemini 4 Carbon: 127.0, $4Claude Fable 5.2: 130.9, $16Claude Fable 5.2Gemini 4 CarbonClaude Fable 5.1Claude Sonnet 5.5Gemini 4 ArgonPrice per 1M tokens, blended (log scale)Overall index
More models in the chart (10)
  • Claude Opus 5.5125.7 · $8
  • GPT-6 Astra119.6 · $20
  • GPT-6.1 Sol117.5 · $4
  • Claude Opus 5117.3 · $10
  • Muse Wonder113.7 · $8
  • Claude Fable 5112.8 · $20
  • Kimi K3.1 Pro112.6 · $4
  • Muse Ember 1.0110.5 · $0.20
  • GPT-5.6 Sol109.3 · $8
  • GPT-6 Sol108.4 · $4

The 15 models with the highest overall index (Gradually Meta-Score) against their API list price of 3 parts input and 1 part output, without discounts or caching. Source: gradually.ai model database.

Chatbots by web traffic

Share of web traffic in September 2026, not user numbers

ChatGPT56.4%
Gemini25%
Claude9.6%
DeepSeek3.4%
Grok2.3%
Copilot1.5%
Perplexity0.9%
Others0.9%

Source: Similarweb (OfficeChai), 09/2026

The development in numbers

How fast new milestones appear, how models grew and where AI is used today.

Milestones per month since 2012

JanFebMarAprMayJun20122013120141120151201612017120181201912020120211112022220231251202423215202534222202631526JulAugSepOctNovDecTotal2012112013120141320151242016122017120181220191202012021320221121182023312133222024122152420251221423202656937
Entries:12-34-56-78-9
Source: gradually.ai AI timeline
gradually.ai

Each cell counts the sourced entries of this timeline in that month. It is not a complete market census.

Parameters (B)
Parameter growth (log)
Best value
OpenAI
Google
NVIDIA
Meta
Mistral AI
Anthropic
Alibaba
Moonshot AI
Others
Source: gradually.ai LLM database
gradually.ai

Many current providers do not publish parameter counts. The dots show only models with a confirmed or estimated count.

AI by country

Outlook: what the forecasts say

None of these years is a fact. A scenario, a community forecast and a survey measure different things.

Forecasts, not events. The four entries describe expectations and scenarios. That is why they are not part of the eras’ timeline.

2027Forecast

Scenario: highly capable AI coding agent

The AI 2027 scenario describes one possible path in which an AI system takes on coding tasks at a high level. It is neither a prediction nor a statistical median.

Basis: AI 2027Source
2030Forecast

2·10²⁹ FLOP training runs become feasible

Epoch AI considers training runs of around 2·10²⁹ FLOP likely technically feasible by 2030 under stated assumptions. Whether a company will actually undertake one remains open. Constraints include power, chips, data, and training time.

Basis: Epoch AISource
2031Forecast

Community forecast: first general AI

On September 25, 2026, the Metaculus community estimates the first general AI in 2031. It is a continuously updated community forecast, not a claim of technical certainty.

Basis: MetaculusSource
2047Forecast

Researcher survey: 50% HLMI

A 2023 survey of 2,778 AI researchers gave 2047 as the median 50% probability for HLMI. HLMI here means machines that can outperform humans at every task.

Basis: Grace et al.Source
Task horizon: from minutes to hours
measured (METR)unreliable (over 16 h)range: doubling every 4.2 to 6.2 months
1 minute1 hour1 work day1 work week1 work monthOct 2026GPT-4 0314GPT-4oClaude 3.5 Sonneto1-previewo1Claude 3.7 Sonneto3GPT-5Gemini 3 ProClaude Opus 4.5Claude Opus 4.6Claude Mythos Preview20242026202820302032
Measured through April 2026 (latest reliable value: 12 h, Claude Opus 4.6). METR flags Claude Mythos Preview at 17.4 h as unreliable because it exceeds 16 hours; it feeds neither the line nor the projection. The task horizon doubles roughly every 6.2 months long-run, lately every 4.2. Logarithmic axis; the projection is not a guarantee.
Source: METR, Time Horizon 1.1 (2026)
gradually.ai
When will general AI arrive? Estimates diverge widely
ScenarioCommunity/marketsResearcher survey
2026203520402045
These estimates measure different things (scenario, community forecast, researcher survey), which is why they span 2027 to 2047. The same researcher survey puts full automation of all jobs at around 2116. The AI 2027 authors have since published longer medians, 2030 for one of them. Forecasts are not guarantees.
Source: AI 2027, Grace et al., Metaculus
gradually.ai

Frequently asked questions about the history of AI

The key questions about the development of AI models and tools.

John McCarthy is generally credited with coining the term. It appears in the 1955 Dartmouth project proposal authored by McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. The 1956 workshop helped establish AI as a research field, decades before ChatGPT.

ChatGPT in November 2022 was the visible catalyst. Its technical foundations came earlier. They included the 2017 Transformer, pretrained language models such as GPT and BERT in 2018, and increasingly capable computing hardware. ChatGPT brought those developments together in a simple conversational interface.

An “intelligence explosion” is a scenario I. J. Good described in 1965. A highly capable machine could design better machines, accelerating further improvements. This is a hypothesis, not an established trend or a measurable metric. The timeline's benchmark charts show task-specific results and do not by themselves demonstrate that such a feedback loop is happening.

There is no universal answer. Open weights can be highly competitive on individual tasks and allow local execution or custom adaptation. Proprietary models may lead on other tasks, tool use, or available infrastructure. Always compare the exact model, its license, the tested configuration, and the source of the result.

Cadence varies by lab and model family. Years can pass between major generations, while smaller variants, updates, and new products may arrive in quick succession. The calendar view counts only sourced entries in this curated timeline, not a complete market statistic.

The timeline covers text and language models, image generators, video and audio AI, and multimodal models. Agents and developer tools appear as a separate category too. You can filter entries by modality, license (open or proprietary), developer, and year.

Release dates, parameter counts, context windows, and benchmark values come from a centrally maintained dataset of official announcements, model cards, and research papers. Each timeline entry links to its source. Benchmark graphics name their data basis because protocol and evaluation date matter for interpretation.

Changelog

The latest updates and improvements to our History of AI
November 12, 2026v1.8

Easier to scan: eras in the page head and at a glance

  • The year bar in the page head now shows which era each year belongs to
  • A new overview sums up the six eras with their key moments and links straight to them
  • The jump bar names the eras and marks the section you are reading
  • Long lists are grouped by year or quarter, and each era shows its key figures
November 9, 2026v1.7

Redesigned: six eras from 1950 to today

  • The history now starts in 1950 and runs through six eras, with the key moments up front as cards
  • A search and a year bar in the page head show right away when a model or company arrived
  • Charts sit in the era they belong to, country data as bars instead of a globe, forecasts in a clearly marked outlook
  • The monthly overview now also counts entries with a day date, and price and performance shows current top models such as GPT-6 and Claude Opus 5.5
September 30, 2026v1.6

New model released September 28

  • Added Claude Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5, succeeding Sonnet 5
Show older versions (6)
September 25, 2026v1.5

Timeline and evidence

  • Added historical foundations and a separate forecast view
  • Activated charts, the globe, and filters on the Astro page
  • Clarified sources, dates, and comparison limits
September 22, 2026v1.4

New models released September 22

  • Added Claude Opus 5.5 as Anthropic's new leading model, succeeding Opus 5
  • Added GPT-6 Sol and GPT-6 Luna as cheaper successors to the GPT-5.6 tiers
September 3, 2026v1.3

New models released September 2

  • Added Muse Spark 1.3 as a new multimodal agentic and coding model
  • Added Qwen 3.8 Max 0902 as a new snapshot with improved coding and agent collaboration
  • Added Gemini 3.8 Flash Cyber as a restricted cybersecurity variant for vetted defenders
  • Added Qwen 3.8 Flash and Qwen 3.8 Flash Next as an API model and open-weight model
August 23, 2026v1.2

July and August leading-model updates

  • Added Claude Sonnet 5, Opus 5, and Kimi K3 to the July and August frontier releases
  • Added Grok 4.5 and Grok 4.6 as new agentic models
  • Added new Gemini Flash variants from July and August to the timeline
  • Added Qwen 3.8 as an open-weights milestone
June 28, 2026v1.1

Prehistory & forecasts

  • Added the deep-learning prehistory: milestones from AlexNet (2012) to AlphaGo and WaveNet (2016)
  • New forecast section with trend extrapolation (METR task horizon, Epoch compute) and sources
  • Forecasts for 2027-2040 as marked entries in the timeline, with uncertainty ranges
June 21, 2026v1.0

Initial release

  • Interactive timeline of generative-AI milestones from 2017 to 2026
  • Filter by modality, license, developer, and year, plus timeline and grid views
  • Stats dashboard with charts on the intelligence explosion (GPQA, SWE-bench, parameters, context)
  • Benchmark scores for older models individually sourced