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.
From 1950 to 2026
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:
139 sourced milestones from 1950 to today, each with a source.
1950-2011
Alan Turing describes the imitation game, later known as the Turing test.
The paper shifts the debate from defining thought to observable behavior.
John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon bring researchers together at Dartmouth.
Their earlier project proposal uses the term 'Artificial Intelligence' and gives the field its name.
Frank Rosenblatt publishes the perceptron, an early trainable neural network. It links mathematical neuron models with a trainable classifier.
Rumelhart, Hinton, and Williams show how errors can be propagated through multiple network layers. The algorithm becomes a practical core of deep learning.
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.
ImageNet publishes a large, hierarchically labelled image dataset. Its later challenge makes progress in image classification measurable and sets the stage for AlexNet.
2012-2016
The AlexNet neural network wins the ImageNet competition by a wide margin, slashing the error rate.
The victory ignites the deep learning revolution.
Ian Goodfellow introduces generative adversarial networks.
Two networks train against each other to produce strikingly realistic images, the basis for many later image generators.
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.
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.
A Google team led by Tomáš Mikolov introduces word2vec. Words become vectors whose distances capture meaning, a foundation of modern language AI.
Google acquires the London AI lab DeepMind. The deal shows that foundational AI research has become strategic for the major platform companies.
Google researchers demonstrate sequence-to-sequence learning with neural networks. The encoder-decoder principle becomes the foundation for machine translation and later the Transformers.
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.
Google releases TensorFlow as open source. An industrial-grade framework makes training neural networks accessible to everyone.
Microsoft Research introduces residual networks. "Skip connections" enable extremely deep networks and win the 2015 ImageNet competition.
DeepMind's WaveNet generates speech directly as raw audio, sounding far more natural than earlier synthesis. The technique later ships in Google products.
2017-2021
A Google team introduces the Transformer architecture, built purely on attention.
It becomes the foundation for GPT, BERT, and every modern language model.
OpenAI releases GPT-3, demonstrating emergent abilities like coding, translation, and reasoning without task-specific training.
GitHub and OpenAI launch Copilot as a technical preview.
It suggests code directly in the editor and brings generative AI into day-to-day software development.
OpenAI publishes the first Generative Pre-trained Transformer model. The approach combines unsupervised pretraining with later adaptation to specific language tasks.
Google introduces BERT. The bidirectionally trained Transformer model shapes search systems and many language-understanding tasks.
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.
OpenAI introduces DALL-E, generating images from text descriptions. Generative AI becomes tangible for a broad audience for the first time.
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.
2022-2023
Stability AI releases Stable Diffusion with open weights.
High-quality image generation suddenly runs free on your own GPU.
OpenAI releases ChatGPT as a conversational interface for a language model.
Its unusually rapid adoption makes generative AI visible to a broad audience.
OpenAI releases GPT-4, which understands images and tops bar and medical exams.
It becomes the benchmark for two years.
Meta releases Llama 2 for free commercial use. A turning point that sparks an entire industry of open-weight models.
Young Paris startup Mistral shows with 7 billion parameters that small open models can beat much larger ones.
20226 entries
DALL-E 2 raises image quality sharply and popularizes diffusion models. The text-to-image race becomes visible to everyone.
With PaLM, Google shows how far pure scaling reaches, reporting strong results on reasoning and multilingual tasks.
Midjourney makes text-to-image publicly accessible through a Discord bot. Its distinct, stylized look quickly becomes a signature.
Meta shows Make-A-Video, generating short clips from text prompts. The results are rough but mark the start of video AI.
OpenAI open-sources Whisper. Accurate transcription across many languages becomes a freely available foundation for audio tools.
Perplexity launches as an AI search engine that answers questions directly with cited sources. The format challenges classic web search.
202319 entries
ElevenLabs opens its beta for speech synthesis and voice cloning. The naturalness of the voices sets a new bar for audio AI.
Microsoft builds a GPT-4 model into Bing. The chatbot draws attention with its quirky replies and opens the search-engine race.
Meta releases LLaMA to researchers; soon after, the weights circulate freely. The open-source community builds its own language models overnight.
Adobe brings Firefly straight into Photoshop and stresses commercially safe training data. Generative AI enters the professional workflow.
Google launches Bard as its own chatbot. The bumpy debut shows how much pressure ChatGPT put on the market leader.
Anthropic introduces the first Claude, positioning itself with "Constitutional AI" as a safety-focused alternative to OpenAI.
With v5, Midjourney produces images barely distinguishable from photos. The first AI images go viral and fool millions.
Runway makes text-to-video publicly usable. The clips are short and flickery, but for the first time anyone can generate moving images.
Anthropic releases Claude 2 with a then-huge 100,000-token context window. Entire books fit into a single prompt.
Stability AI follows up with SDXL, markedly improving composition and detail in open image models.
Meta specializes Llama 2 for code. Open-source development gains a freely usable programming assistant.
Abu Dhabi's Technology Innovation Institute releases Falcon 180B, briefly topping the open-model leaderboard.
DALL-E 3 moves into ChatGPT and follows complex instructions far more closely. Image generation becomes a conversation.
At its first DevDay, OpenAI shows GPT-4 Turbo with a 128,000-token context and custom configurable GPTs. The agent idea takes shape.
Elon Musk's xAI introduces Grok, tied closely to X and deliberately cheeky in tone. A few months later the weights are opened.
Startup Pika brings accessible video generation and raises substantial venture capital. The video-AI race speeds up.
Google introduces Gemini, built from the ground up for text, image, audio, and video. It rivals GPT-4 across many benchmarks.
With v6, Midjourney reliably renders text inside images for the first time and pushes photographic quality further.
Mistral releases Mixtral, showing how a mixture-of-experts model delivers strong performance at lower cost per query.
A large context window describes the maximum input length, not the quality across the whole context.
2024-2025
OpenAI shows Sora demos with photorealistic clips up to a minute long.
The quality is years ahead of rivals and sparks a debate.
Anthropic releases Claude 3 (Opus, Sonnet, Haiku), beating GPT-4 on many benchmarks.
For the first time, OpenAI is no longer unchallenged.
Meta releases Llama 3.1 405B, the first open model to keep pace with the best proprietary systems.
Black Forest Labs, founded by Stable Diffusion's inventors, releases FLUX.1 and sets a new standard for open image models.
DeepSeek trains an open 671-billion model for about $5.6 million, calling into question the entire industry's cost assumptions.
Google releases Gemini 2.5 Pro, topping many leaderboards clearly for the first time.
The company is back in the race for the best model.
202419 entries
Google shows Gemini 1.5 Pro with a one-million-token context window. Whole codebases or hours of video fit into a single prompt.
Cognition introduces Devin, an agent that plans and executes whole coding tasks on its own. The vision of autonomous coding agents becomes tangible.
Suno turns a short prompt into finished songs with vocals and instruments. AI music becomes mainstream overnight.
Meta releases Llama 3 with performance close to proprietary models. Open AI becomes truly competitive.
Udio launches with especially clean audio quality and challenges Suno. The AI music market becomes a duel.
OpenAI ships GPT-4o, processing text, image, and audio in real time. Natural spoken conversation with AI becomes reality.
Anthropic releases Claude 3.5 Sonnet, beating Claude 3 Opus at lower cost, and introduces "Artifacts" for interactive output.
Chinese company Kuaishou launches Kling, showing that top-tier video models do not come only from the West.
Luma makes high-quality video generation freely accessible. Demand is so high the servers buckle at times.
Runway answers Sora with Gen-3 Alpha, delivering markedly more coherent, cinematic video.
Stability AI releases Stable Diffusion 3 with better prompt adherence and typography. Licensing questions, however, spark debate.
xAI releases Grok 2 with built-in image generation, moving close to the top tier of chatbots.
With NotebookLM, Google turns your own documents into surprisingly natural podcast dialogues. The feature becomes a viral hit.
OpenAI releases o1, which "thinks" before answering. Spending more compute at inference time becomes a new source of capability gains.
Anthropic gives Claude 3.5 Sonnet the ability to operate mouse and keyboard. The agent can use software like a human.
Google ships Gemini 2.0 Flash with native image and audio output, focusing on agentic applications.
With Llama 3.3, Meta shows a 70-billion model nearly matching the five-times-larger 405B model.
OpenAI finally makes Sora publicly available. By now rivals have caught up, and the February lead has melted away.
Google shows Veo 2 with a better grasp of physics and camera motion, positioning directly against Sora.
202522 entries
AI code editor Cursor races to hundreds of millions in revenue. Coding with an AI agent inside the editor becomes normal for many teams.
Chinese startup DeepSeek releases an open reasoning model on par with o1. The news sends AI stocks tumbling and reshapes the global landscape.
OpenAI launches Operator, an agent that operates websites on its own and handles tasks like bookings. Agentic AI leaves the lab.
Anthropic combines fast answers and visible thinking in one model. Users control how long Claude "thinks" about a task.
Anthropic introduces Claude Code, an agent that handles whole tasks across many files in the terminal. It becomes the template for agentic coding.
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.
xAI trains Grok 3 on a massive GPU cluster with ten times the compute of Grok 2, closing in on the leaders.
OpenAI builds native image generation into GPT-4o. Millions turn photos into anime styles, and the servers hit their limits for days.
Meta introduces Llama 4 Scout and Maverick as open MoE models, with Scout offering a ten-million-token context window.
OpenAI releases o3 and o4-mini, which use tools like web search and code on their own while thinking. Agentic reasoning becomes standard.
Anthropic releases Claude Opus 4 and Sonnet 4, able to work on code autonomously for hours. Claude takes the lead in programming.
Veo 3 generates clips with matching sound and dialogue for the first time. Generated video becomes hard to tell from real footage.
xAI releases Grok 4, reporting top scores on several reasoning benchmarks. The Colossus GPU cluster pays off.
OpenAI ships GPT-5, automatically blending fast answers and deep reasoning based on the task. Manual model switching disappears.
Google's image model nicknamed "Nano Banana" edits photos consistently across multiple steps. Targeted editing instead of regenerating becomes the norm.
Anthropic releases Claude Sonnet 4.5, which by its own account can work autonomously on complex tasks for over 30 hours.
OpenAI ships Sora 2 with synchronized audio and a dedicated app for AI videos. Cameos let users place themselves into generated clips.
Anthropic releases Claude Opus 4.5, cementing its position on agentic and coding tasks.
DeepSeek updates its open MoE model and keeps the gap to proprietary leaders small, at a fraction of the cost.
Google releases Gemini 3 Pro with markedly improved reasoning and multimodal understanding, setting new highs.
OpenAI follows up with GPT-5.2, improving speed and reasoning over the original GPT-5.
Paris-based Mistral releases a 675-billion MoE model, keeping the European flag flying in the AI race.
GPQA Diamond tests graduate-level science knowledge. The values come from different configurations and evaluations, not a normalized ranking.
SWE-bench Verified starts in August 2024. Scaffolds and compute budgets differ per result.
2026
Anthropic releases Opus 4.6 and Sonnet 4.6 with a 1-million-token context.
Opus 4.6 hits 76% on the MRCR v2 benchmark and coordinates "agent teams".
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.
DeepSeek releases an open MoE model with 1.6 trillion parameters (49 billion active) under an MIT license.
Open models reach trillion scale.
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.
Moonshot AI initially releases Kimi K3 as an API model.
The frontier model combines text with native visual capabilities.
Alibaba publishes the Qwen 3.8 weights. The release marks another open-weights milestone for frontier models.
January to March2 entries
OpenAI unifies frontier coding and professional reasoning in a single model for the first time. It runs about 25% faster than its predecessor.
OpenAI ships GPT-5.4, hitting 92.8% on GPQA Diamond with gains in knowledge and multilingual tasks.
April to June11 entries
Anthropic releases Opus 4.7 with a new tokenizer, adaptive thinking, task budgets for agents, and higher-resolution vision. SWE-bench Verified: 87.6%.
Moonshot AI releases Kimi K2.6 with 1 trillion parameters, native INT4 quantization, and a 262,000-token context under a modified MIT license.
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.
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%.
At Google I/O 2026, Gemini 3.5 Flash arrives with four times faster output and strong scores on agentic benchmarks.
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.
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.
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.
Shanghai lab MiniMax ships M3, an API model with a 1-million-token context. Another Chinese provider pushes toward the frontier.
Mistral AI updates its Medium line with a coding-focused API model. The parameter count stays undisclosed.
Anthropic releases Claude Sonnet 5 for complex agentic tasks. The new generation extends the line with a frontier model for professional workflows.
July to September18 entries
xAI advances the Grok line with a new agentic model. Grok 4.5 extends the family for complex tasks involving tools.
Google further expands the Flash family. Gemini 3.6 Flash targets fast multimodal applications.
Anthropic positions Claude Opus 5 as a new model for demanding reasoning. It leads the Claude 5 generation in the most complex professional workflows.
Moonshot AI releases the Kimi K3 weights. This makes a frontier model with native visual capabilities openly available.
xAI releases Grok 4.6 as its next agentic model version. The Grok line continues to advance at a rapid pace.
Google updates its fast multimodal Flash line again. Gemini 3.7 Flash continues the dense summer release cadence.
Z.ai releases GLM-5.3 as a new API model for agentic tasks. The model extends the GLM 5 line for professional applications.
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.
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.
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.
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.
Google introduces Gemini 3.8 Flash Cyber for vulnerability discovery and automated patching. Access remains limited to vetted defenders in the Fairwind program.
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.
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.
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.
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.
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.
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.
most attractive area: stronger and cheaper than half of the models shownbest intelligence for the price
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.
Share of web traffic in September 2026, not user numbers
Source: Similarweb (OfficeChai), 09/2026
Each cell counts the sourced entries of this timeline in that month. It is not a complete market census.
Many current providers do not publish parameter counts. The dots show only models with a confirmed or estimated count.
Gradually LLM selection by developer location
238 of 303 entries in the LLM selection have a country mapping. The remaining entries are omitted here. Countries without a value have no mapped entries; this does not measure their model production.
Source: Gradually LLM selection (Nov 2026)
Share of ChatGPT desktop traffic by country
The publicly listed top five countries. Similarweb estimates desktop visits to chatgpt.com; app use and mobile browsers are outside this dataset. Other countries are not listed individually.
Private AI investment, US and China
This selection shows the United States and China. Other countries have no value in this dataset; missing values do not mean zero investment.
ChatGPT availability by country
Forecasts, not events. The four entries describe expectations and scenarios. That is why they are not part of the eras’ timeline.
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.
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.
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.
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.