AI History Timeline

Every major AI milestone from the Transformer paper to today. 68 dated events across 11 years, plus 3 expected releases.

Dated entries verified against primary announcements on September 8, 2026 · newest entry: Gemini 4 Argon (Sep 30, 2026)

2027

  1. Expected

    Recursive Self-Improvement (Expected)

    Per xAI co-founder Jimmy Ba and others, AI systems autonomously upgrading their own capabilities expected to emerge in H1 2027. Considered by many the threshold of meaningful AGI progress.

    Industry

2026

  1. Expected

    First True Multi-Agent Frameworks

    Production-ready frameworks for orchestrating teams of specialized AI agents expected to mature by end of 2026 — collaborating across organizations, not just within single workflows.

    Industry

  2. Expected

    Claude Haiku 5.5 (Expected)

    The last model of the Claude 5.5 family. Anthropic has said twice that it will follow "in the coming weeks", at the Opus 5.5 launch on 22 September and again at Sonnet 5.5's on 28 September, without giving a date.

    Anthropic

  3. Model Release

    Gemini 4 Argon

    Google's first Gemini 4 model, with a 1M-token output limit and an introductory $2/$10 that later doubles. It leads Google's own table on 13 of 19 tests and ties GPT-6 Astra on the Artificial Analysis index for 61% of the cost per task, but reaches cyber defenders in the Fairwind Program first, with no date for paid API customers. Gemini 3.5 Pro never shipped.

    Google

  4. Model Release

    GPT-6.1 Sol

    A week after GPT-6 Sol and a day after OpenAI scrapped GPT-6.1 Astra, the DevDay model keeps Sol's $2/$10 and halves cached input to $0.10. On OpenAI's charts it edges Astra on coding for a seventh of the cost per task, but trails Astra and Claude Opus 5.5 on business automation.

    OpenAI

  5. Model Release

    Claude Sonnet 5.5

    Near-Opus scores at Sonnet 5's unchanged $2/$10: two points behind Opus 5.5 on the Artificial Analysis Intelligence Index, and the first Sonnet to ship with cyber safeguards. The catch is maximum effort, where it costs more per task than Opus 5.5 itself.

    Anthropic

  6. Model Release

    Claude Opus 5.5

    A cheaper model takes first place: $4/$20, a fifth under Opus 5, and the top score on the Artificial Analysis Intelligence Index — ahead of Claude Fable 5.1, which costs more than twice as much per token. Thinking can no longer be switched off.

    Anthropic

  7. Model Release

    GPT-6 Sol and GPT-6 Luna

    The rest of the GPT-6 family lands the same day, at half the GPT-5.6 rates: Sol at $2/$10 and Luna at $0.10/$0.50, a 100× spread from Luna to Astra. Sol, the GPT-5.6 flagship, is now the middle tier, and free ChatGPT users get Luna on desktop.

    OpenAI

  8. Model Release

    Grok 4.7

    A larger base model and a longer reinforcement-learning run, at exactly Grok 4.6's $2/$6. It beats its predecessor on every benchmark SpaceXAI published and trails Claude Fable 5.1 on most, at a fifth of the input price and an eighth of the output.

    SpaceXAI

  9. Model Release

    Meta Muse Spark 1.3

    Meta's fourth frontier model in five months takes the top spot on DeepSWE and near-saturates million-token retrieval. A "contributor" endpoint undercuts every rival at ~$0.10 in and $0.20 out — the consideration being that Meta may train on your prompts.

    Meta

  10. Model Release

    GPT-6 Astra

    OpenAI's largest training run ever, and the first model it has rated Critical for cyber capability. Drives a computer at superhuman speed, ships a 1,050,000-token context window, and raises list price to $10/$50 while using far fewer tokens per finished task.

    OpenAI

  11. Model Release

    Claude Fable 5.1

    A point release with the largest single benchmark jump Anthropic has published: double the Terminal-Bench-Science score, cache reads cut 75% to $0.25 per million, and headline pricing unchanged at $10/$50. Ships beside Mythos 5.1.

    Anthropic

  12. Company

    Anthropic Targets a Record IPO

    Anthropic signals an offering larger than SpaceX's record listing, capping a year in which frontier-lab valuations outran every prior tech cycle.

    Anthropic

  13. Company

    Stripe Buys OpenRouter for $7B

    Five times its May price. Payments infrastructure moves to own the model-routing layer as inference becomes a metered commodity.

    Stripe

  14. Regulation

    First AI Music Ruling in Europe

    Suno loses to German collecting society GEMA — Europe's first court ruling on AI-generated music and training data.

    GEMA/Suno

  15. Regulation

    EU AI Act Hits Its Big Deadline

    The headline compliance date arrives after Brussels quietly moved most substantive obligations to 2027 — enforcement, not the statute, becomes the open question.

    European Union

  16. Milestone

    OpenAI Names Its Next Family Astra

    Announced alongside a claim that the models solved ten decade-old open math problems — and, days later, that OpenAI cannot rule out critical cyber risk in them.

    OpenAI

  17. Model Release

    Claude Opus 5

    1M context, a new "xhigh" reasoning mode, and the first Claude to make the effort dial a headline feature rather than a setting.

    Anthropic

  18. Model Release

    Kimi K3 — Largest Open-Weight Model Ever

    A 2.8-trillion-parameter open-weight model that beats Fable 5 at frontend code; the full weights go free eight days later.

    Moonshot AI

  19. Regulation

    Export Ban on Fable 5 and Mythos 5 Lifted

    Washington reverses course after 76 security veterans argue the ban disarmed defenders more than attackers. Both models return to market.

    United States

  20. Model Release

    Claude Sonnet 5

    The mid-tier lands close enough to Opus 4.8 to make the price gap the main reason to choose between them.

    Anthropic

  21. Model Release

    GPT-5.6 Ships as Sol, Terra and Luna

    OpenAI splits a release into three separately priced models instead of one, ending up with a 25× price spread across the family.

    OpenAI

  22. Company

    SpaceX Buys Cursor for $60B

    An all-stock merger that hands xAI a coding powerhouse, and the largest acquisition yet in AI developer tooling.

    SpaceX/xAI

  23. Regulation

    U.S. Export Controls Pull Fable 5 and Mythos 5

    Anthropic withdraws both models worldwide to comply with a U.S. export-control directive — the first time a frontier model is pulled from the global market on national-security grounds.

    United States

  24. Model Release

    Claude Fable 5

    Anthropic's most powerful public model yet, shipped with the cyber capabilities deliberately sanded off after months of leak-driven speculation.

    Anthropic

  25. Milestone

    Apple Puts Gemini Behind Siri

    At WWDC, Siri becomes Gemini-powered and Claude becomes a first-party iPhone option — Apple concedes the model layer to outside labs.

    Apple

  26. Company

    Anthropic Files Confidentially for an IPO

    Beats OpenAI to the SEC at a reported $965B valuation, opening the race to be the first frontier lab on the public markets.

    Anthropic

  27. Milestone

    Copilot Moves to Usage-Based Billing

    Premium requests out, AI credits in — the first major coding assistant to price by consumption instead of per seat, and the template the rest of the industry follows.

    GitHub

  28. Model Release

    Claude Opus 4.8

    Ships 41 days after 4.7. Dynamic Workflows plan and run hundreds of parallel subagents per session, a user-facing effort dial, and a 2.5× faster mode at 3× lower cost — plus its "most honest" self-review yet, all at 4.7's price.

    Anthropic

  29. Milestone

    Cursor SDK Released

    Build custom agents using the same runtime that powers Cursor. /multitask enables async subagents in parallel.

    Cursor

  30. Milestone

    Sora App Shutdown

    OpenAI closes the standalone Sora app and website with no successor product — Sora 2 is not folded into ChatGPT — and schedules the Sora API to follow on September 24, 2026. End of an era for consumer video AI.

    OpenAI

  31. Model Release

    GPT-5.5 Launched

    OpenAI's most capable and intuitive model. Strong gains in agentic coding and computer use. Outperforms Gemini 3.1 Pro and Claude Opus 4.5 on most benchmarks.

    OpenAI

  32. Model Release

    Claude Opus 4.7

    First Claude with high-resolution vision (2576px). 1M context, 128K max output. New tokenizer.

    Anthropic

  33. Milestone

    Claude Code Desktop Redesign

    Complete redesign with multi-session sidebar, integrated terminal, file editor, diff viewer. Sets new standard for AI coding tools.

    Anthropic

  34. Model Release

    Gemini 3.1 Ultra

    Google's highest tier breaks reasoning records with 94.3% on GPQA Diamond, aimed at enterprise agentic workflows.

    Google

  35. Model Release

    Llama 5 (Open)

    Meta's 600B open-weight release, claiming frontier-level performance and keeping the open-model frontier within reach of the closed labs.

    Meta

  36. Model Release

    Claude Mythos 5

    The first 10-trillion-parameter model — a step change in capability, and the model that later triggered U.S. export controls.

    Anthropic

  37. Model Release

    Gemma 4 (Open)

    Open-weights model purpose-built for advanced reasoning and agentic workflows.

    Google

  38. Model Release

    Suno v5.5 — Custom Voices

    Most-requested feature ships: subscribers can upload their own singing voice and use it in AI-generated songs.

    Suno

  39. Model Release

    Midjourney V8 Alpha

    5x faster generation, native 2K resolution, improved text rendering. Major upgrade since V7.

    Midjourney

  40. Model Release

    Gemini 3.1 Flash Lite

    Ultra-fast tier for high-volume use cases. Aggressive pricing.

    Google

  41. Model Release

    Nano Banana 2 (Image)

    Best-in-class image generation from Google, built on Gemini 3.1 Flash Image. Strong typography.

    Google

  42. Model Release

    Gemini 3.1 Pro

    Largest production context window: 1M tokens. Improved reasoning across the board.

    Google

  43. Model Release

    Claude Sonnet 4.6

    First Sonnet to beat the prior generation's Opus on coding benchmarks. Best price/perf in the market.

    Anthropic

2025

  1. Company

    Anthropic-Warner Music Partnership

    First major music label partnership with an AI company — sets terms for AI music licensing.

    Anthropic

  2. Regulation

    EU AI Act Fully In Force

    World's first comprehensive AI regulation reaches full enforcement, including foundation model disclosures.

    European Parliament

  3. Model Release

    GPT-5 Launched

    First in the GPT-5 family. Sets new bar for agentic capabilities and computer use.

    OpenAI

  4. Model Release

    Claude Opus 4 Released

    Major leap in coding and agentic capabilities. Native tool use, dramatically improved Claude Code experience.

    Anthropic

  5. Model Release

    Llama 4 Released (Open)

    Mixture-of-Experts open-weight model competitive with closed frontier models.

    Meta

  6. Model Release

    Claude Sonnet 3.7 Released

    First Sonnet with extended thinking. Pushed coding capabilities significantly forward.

    Anthropic

  7. Model Release

    Gemini 2.0 + Deep Research

    Native multimodal with audio output and Deep Research agent. First flagship with 1M token context window.

    Google

2024

  1. Research

    Anthropic Releases MCP

    Model Context Protocol launched — the open standard for connecting LLMs to tools and data sources.

    Anthropic

  2. Model Release

    o1 (Reasoning) Released

    First widely-deployed "reasoning model" — used test-time compute to dramatically improve math/logic.

    OpenAI

  3. Model Release

    FLUX.1 Released

    First open-weight image model competitive with Midjourney for photorealism.

    Black Forest Labs

  4. Model Release

    GPT-4o Released

    Truly multimodal flagship — text, audio, vision in one model. Voice Mode brought natural AI conversation to ChatGPT.

    OpenAI

  5. Model Release

    Claude 3 Family Released

    Opus, Sonnet, Haiku tiers. Opus briefly took the top of every benchmark.

    Anthropic

  6. Milestone

    Sora Announced

    First major text-to-video model. Stunned the world with photorealistic minute-long clips.

    OpenAI

2023

  1. Model Release

    Gemini 1.0 Launched

    Google enters the LLM race with Gemini, available in Ultra/Pro/Nano variants.

    Google

  2. Model Release

    Llama 2 Open Source

    Meta releases Llama 2 with commercial use allowed. Kickstarted the open-source LLM ecosystem.

    Meta

  3. Model Release

    GPT-4 Released

    Significant capability jump. Multimodal (images + text), advanced reasoning, code generation.

    OpenAI

  4. Model Release

    Claude 1 Released

    Anthropic launches its first public AI assistant, focused on safety and helpfulness.

    Anthropic

2022

  1. Milestone

    ChatGPT Launched

    The moment AI went mainstream. ChatGPT reached 1M users in 5 days and changed everything.

    OpenAI

  2. Model Release

    Stable Diffusion Released (Open Source)

    First open-source state-of-the-art image generator. Democratized AI image creation overnight.

    Stability AI

  3. Model Release

    DALL-E 2 Released

    Breakthrough text-to-image quality. Closed beta access — public couldn't use it freely yet.

    OpenAI

2021

  1. Company

    Anthropic Founded

    Dario and Daniela Amodei plus former OpenAI researchers found Anthropic, focused on AI safety.

    Anthropic

  2. Milestone

    GitHub Copilot Launched

    First mainstream AI coding assistant, powered by Codex (a GPT-3 variant). Changed how developers work.

    GitHub/OpenAI

2020

  1. Model Release

    GPT-3 Released

    175B parameter GPT-3 launched via API. Sparked massive developer interest and the modern LLM era.

    OpenAI

2019

  1. Model Release

    GPT-2 Released (Cautiously)

    GPT-2 (1.5B parameters) released gradually due to "misuse concerns" — first large public discussion of AI safety.

    OpenAI

2018

  1. Model Release

    BERT Released

    Google publishes BERT, revolutionizing NLP for understanding tasks (search, classification).

    Google

  2. Model Release

    GPT-1 Released

    OpenAI releases the first Generative Pre-trained Transformer (117M parameters), launching the GPT family.

    OpenAI

2017

  1. Research

    Transformer Architecture Published

    Vaswani et al. publish "Attention Is All You Need" — the paper that introduced the Transformer architecture, the foundation of every modern LLM.

    Google

How this timeline is built

This is a curated record, not a feed. An event earns a place when it is a dated, primary-source milestone that changed what people could do with AI or how the industry is organised: a model release, a product that shipped to the public, a regulation taking effect, a founding, an acquisition, a market-shaping deal. Rumours, benchmark bumps and funding rounds stay in the news feed; the timeline is what still matters once the week's noise is gone. Each entry carries the date its primary announcement gives — a full date where one exists, a month where the source only offers that — and the company behind it.

How to read it

The colour tag on each card is the kind of event. Model release covers new models and major versions that have shipped, and those cards also carry a green frame; Milestoneis a product or capability moment, like ChatGPT's launch, a model demo before it ships or Copilot moving to usage-based billing; Company is corporate — foundings, acquisitions, IPO filings; Regulation is law and export control; Research is a paper or protocol that the field built on. Dashed Expectedcards are announced or widely anticipated but not shipped: they are dated by month from vendor roadmaps and on-record statements, and they move into the dated history with their real date once they happen. Four stale predictions once sat on this page as “upcoming” months after launching, which is why the expected block is now reviewed with the rest.

Verification and corrections

The dated entries were last checked against their primary announcements on September 8, 2026. When an entry turns out to be wrong it is corrected in place and the correction is noted here rather than quietly overwritten: on September 8, 2026 the Sora App Shutdown entry was corrected — an earlier version said Sora's video generation had moved into ChatGPT Pro, but OpenAI's discontinuation notice names no successor and no ChatGPT integration, and schedules the API to follow on September 24, 2026. For the context behind the entries, our plain-English guide to agentic AI explains the shift the 2025–2026 releases describe, the Claude and ChatGPT guides cover the two product lines that dominate the recent years, and the leaderboard shows how the models on this page rank today.

Frequently Asked Questions

How are events chosen for this timeline?

An entry has to be a dated, primary-source event — a model release, a shipped product, a regulation taking effect, a company founding or acquisition — that changed what people could do with AI or how the industry works. Rumours, benchmarks and funding rounds are left to the news feed.

What does the dashed 'Expected' label mean?

Those entries are announced or widely expected but have not shipped. They are drawn from vendor roadmaps and on-record statements, dated by month, and moved into the dated history with their real date once they happen. If a prediction is abandoned, the entry is removed rather than left standing.

How up to date is the timeline?

The dated entries were last checked against their primary announcements on September 8, 2026. New milestones are usually added within days of shipping; the newest entry shows the most recent one.

Where does the 'Sora App Shutdown' entry get its facts?

From OpenAI's own discontinuation notice: the Sora app and website closed on April 26, 2026 with no successor product, and the API follows on September 24, 2026. An earlier version of this entry said Sora had moved into ChatGPT Pro; that was wrong and was corrected on September 8, 2026.

Missing an event? Email usand we'll add it.