Timeline¶
Last checked November 8, 2026. It covers January 2017 to early November 2026.
This page lists dated events in AI-accelerated software engineering: key papers, model launches, tools, standards and incidents. Each row links to a primary source where one exists, or to reputable reporting where it doesn't.
How to read it:
- Dates are first public availability in US time. API dates can lag app launches by days. Where they differ, both are noted. "YYYY-MM" means only the month is certain.
- Model details (prices, context windows, IDs) live in module 03. This page gives just enough to place each event.
- "Reported" marks events confirmed only by secondary sources.
- The field moves monthly. If you find an error, fix it here and in module 03.
2017–2019: the transformer and the first GPTs¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2017-01-23 | Sparsely-gated mixture-of-experts layer paper | Origin of the MoE design behind most 2026 open models | arXiv |
| 2017-06-12 | "Attention Is All You Need" (Vaswani et al.) | Introduced the transformer, the architecture behind every model here | arXiv |
| 2018-06-11 | OpenAI GPT-1 paper ("Improving Language Understanding by Generative Pre-Training") | Showed pretraining plus fine-tuning works for language | OpenAI (PDF) |
| 2018-10-11 | Google's BERT | Made pretrained transformers the default in NLP | arXiv |
| 2019-02-14 | OpenAI GPT-2, initially withheld over misuse concerns | First major "too dangerous to release" debate about model weights | OpenAI |
2020–2021: scaling, and code models arrive¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2020-01-23 | Scaling laws for neural language models (Kaplan et al.) | Showed loss falls predictably with model size, data and compute | arXiv |
| 2020-05-28 | GPT-3 paper, "Language Models are Few-Shot Learners" | Prompting alone could do new tasks; start of the API era | arXiv |
| 2021-01-11 | Switch Transformers | Scaled MoE to a trillion parameters | arXiv |
| 2021-06-17 | LoRA paper | Cheap fine-tuning that later made open-model customization practical | arXiv |
| 2021-06-29 | GitHub Copilot technical preview | First mass-market AI pair programmer (autocomplete in the IDE) | GitHub |
| 2021-07-07 | OpenAI Codex paper and HumanEval benchmark | The model behind Copilot; HumanEval became the first standard coding benchmark | arXiv |
2022: instruction tuning, reasoning prompts and ChatGPT¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2022-01-28 | Chain-of-thought prompting paper | Worked examples that spell out intermediate steps improved reasoning, the seed of reasoning models | arXiv |
| 2022-03-04 | InstructGPT (RLHF) paper | Human-feedback training made models follow instructions | arXiv |
| 2022-03-29 | Chinchilla scaling paper | Data matters as much as size; reshaped how models are trained | arXiv |
| 2022-10-06 | ReAct paper (reason plus act) | The pattern behind tool-using agents | arXiv |
| 2022-11-30 | ChatGPT launches | Brought LLMs to the public; fastest-growing consumer app of its time | OpenAI |
| 2022-12-15 | Anthropic's Constitutional AI paper | Training with written principles; basis of Claude's alignment approach | arXiv |
2023: GPT-4, Claude, Llama and open weights¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2023-02-09 | Toolformer paper | Models can learn when to call tools | arXiv |
| 2023-02-24 | Meta's LLaMA | Leaked, then widely used; started the open-weight boom | arXiv |
| 2023-03-14 | GPT-4 | Big jump in reasoning and coding; set the frontier for a year | OpenAI |
| 2023-03-14 | Anthropic launches Claude | Second major closed-frontier lab enters | Anthropic |
| 2023-05-23 | QLoRA paper | Fine-tune large models on one GPU | arXiv |
| 2023-07-11 | Claude 2 | 100K context; Claude becomes a coding option | Anthropic |
| 2023-07-18 | Llama 2, licensed for commercial use | Made open weights viable for companies; OSI disputed "open source" | Meta, OSI |
| 2023-09-12 | vLLM / PagedAttention paper | Efficient serving; vLLM became the standard open-model server | arXiv |
| 2023-10-10 | SWE-bench | Real GitHub issues as a coding benchmark; later the yardstick for agents | arXiv |
| 2023-10-31 | Papers show LoRA cheaply removes Llama 2's safety training | Core evidence in the open-weight safety debate | arXiv |
| 2023-12-06 | Google Gemini 1.0 | Google's native multimodal family | |
| 2023-12-11 | Mistral's Mixtral 8x7B (open MoE) | Showed open MoE models could rival larger dense ones | Mistral |
2024: long context, agents and MCP¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2024-03-04 | Claude 3 family (Haiku, Sonnet, Opus) | Introduced the three-tier naming still used in 2026 | Anthropic |
| 2024-03-12 | Cognition's Devin, "AI software engineer" demo | Popularized the autonomous coding agent idea, and the hype debate | Cognition |
| 2024-04-18 | Llama 3 | Strong open weights from Meta | Meta |
| 2024-05-13 | GPT-4o | Fast, cheap multimodal flagship; later the center of the 2025 sycophancy and GPT-5 backlash stories | OpenAI |
| 2024-06-20 | Claude 3.5 Sonnet and Artifacts | Became developers' favorite coding model for a year | Anthropic |
| 2024-07-23 | Llama 3.1 405B; Zuckerberg's "Open Source AI Is the Path Forward" | First frontier-scale open model; Meta's open strategy at its peak | Meta, Meta |
| 2024-07-30 | NTIA report on open model weights | US government recommended monitoring, not restricting, open weights | NTIA |
| 2024-08-13 | SWE-bench Verified | Human-checked subset that became the headline coding benchmark | OpenAI |
| 2024-09-12 | OpenAI o1-preview | First widely available reasoning model; thinking before answering | OpenAI |
| 2024-10-22 | Claude computer use and upgraded Claude 3.5 Sonnet | Models controlling GUIs; big jump on SWE-bench | Anthropic |
| 2024-10-28 | OSI Open Source AI Definition 1.0 | Defined "open source AI": weights, code and data information | OSI |
| 2024-11-04 | Epoch AI: open models lag the frontier by about a year | First careful estimate of the open-closed gap | Epoch AI |
| 2024-11-25 | Anthropic introduces the Model Context Protocol (MCP) | Open standard for connecting models to tools and data | Anthropic |
| 2024-12-26 | DeepSeek-V3 | Frontier-class open MoE trained cheaply; set up the R1 moment | arXiv |
2025: reasoning, coding agents and vibe coding¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2025-01-20 | DeepSeek-R1 (open weights, MIT) | Open reasoning model near o1; shook markets; start of Chinese open-weight leadership | arXiv, DeepSeek |
| 2025-02-02 | Andrej Karpathy coins "vibe coding" | Named the practice of "forgetting the code exists" and accepting AI output | X |
| 2025-02-24 | Claude 3.7 Sonnet and Claude Code (research preview) | Anthropic's terminal coding agent debuts | Anthropic |
| 2025-02-27 | OpenAI GPT-4.5 | Largest non-reasoning GPT; showed limits of pure scale | OpenAI |
| 2025-03-18 | METR "Measuring AI Ability to Complete Long Software Tasks" | Time-horizon metric: task length agents can do doubles about every 7 months | arXiv |
| 2025-03-25 | Gemini 2.5 Pro | Google takes the lead on many leaderboards | |
| 2025-04-05 | Llama 4 Scout and Maverick | Disappointing reception; Arena used an unreleased "experimental" variant | Meta, TechCrunch |
| 2025-04-14 | GPT-4.1 (API) | 1M-token context; tuned for coding and instruction following | OpenAI |
| 2025-04-16 | OpenAI o3 and o4-mini; Codex CLI open-sourced | OpenAI's terminal coding agent | OpenAI, GitHub |
| 2025-04-29 | OpenAI rolls back a sycophantic GPT-4o update | A model update can change personality in harmful ways overnight | OpenAI |
| 2025-04-29 | "The Leaderboard Illusion" paper | Documented private Arena testing (27 Llama 4 variants) | arXiv |
| 2025-04-29 | Qwen3 family (open, Apache 2.0) | Qwen becomes the most fine-tuned open base | Qwen |
| 2025-05-16 | OpenAI Codex cloud agent (research preview) | Parallel coding tasks in cloud sandboxes | OpenAI |
| 2025-05-19 | GitHub Copilot coding agent | Assign an issue, get a pull request | GitHub |
| 2025-05-22 | Claude Opus 4 and Sonnet 4; Claude Code generally available | Claude Code becomes a mainstream tool | Anthropic |
| 2025-06-12 | SEAL, "Self-Adapting Language Models" (MIT) | Models generate fine-tuning data and update directives; an optimizer updates the weights | arXiv |
| 2025-06-19 | Tobi Lütke popularizes "context engineering" (Karpathy follows on Jun 25) | Shift from prompt wording to managing everything in the context | X (Lütke), X (Karpathy) |
| 2025-06-25 | Google Gemini CLI (open source) | Third major terminal agent, with a free tier | |
| 2025-07-09 | xAI Grok 4 | xAI's first model at the top of several benchmarks | xAI |
| 2025-07-10 | METR randomized trial: experienced developers 19% slower with early-2025 AI tools | The most cited skeptical productivity result | arXiv |
| 2025-07-11 | Moonshot Kimi K2 (1T-parameter open MoE) | Open model built for agentic coding | Moonshot |
| 2025-07-22 | Qwen3-Coder | Open coding model with its own CLI | Qwen |
| 2025-07-23 | America's AI Action Plan encourages open models | US policy turns explicitly pro-open-weight | White House (PDF) |
| 2025-07-28 | Z.ai GLM-4.5 | GLM becomes a leading open coding model | Z.ai |
| 2025-08-04 | Nathan Lambert launches the ATOM Project | Argued the US had lost open-model leadership to China | Interconnects |
| 2025-08-05 | OpenAI gpt-oss-120b and 20b (Apache 2.0) | OpenAI's first open-weight language models since GPT-2 | OpenAI, Hugging Face |
| 2025-08-05 | Claude Opus 4.1 | Incremental coding upgrade | Anthropic |
| 2025-08-07 | GPT-5 launches; router outage, "chart crime," GPT-4o backlash | A rocky launch; 4o restored for paid users within a day | OpenAI, TechCrunch |
| 2025-08 | AGENTS.md published as a shared format for agent instructions | One file for many agents instead of CLAUDE.md, .cursorrules and others | agents.md |
| 2025-09-15 | GPT-5-Codex | Coding-tuned GPT-5 for Codex (API Sep 23) | OpenAI |
| 2025-09-17 | Anthropic postmortem on three infrastructure bugs | "We never reduce model quality due to demand"; quality dips were bugs | Anthropic |
| 2025-09-29 | Claude Sonnet 4.5 | Long-running agent work; new top coding model at launch | Anthropic |
| 2025-09-30 | NIST CAISI evaluation of DeepSeek models | Found performance gaps and higher agent-hijacking rates | NIST |
| 2025-10-15 | Claude Haiku 4.5 | Small model near earlier Sonnet quality | Anthropic |
| 2025-10-16 | Anthropic Agent Skills | Folders of instructions and scripts loaded on demand | Anthropic |
| 2025-10-29 | Cursor 2.0 and its Composer model | Leading AI IDE ships its own coding model and multi-agent UI | Cursor |
| 2025-11-06 | Kimi K2 Thinking | Open reasoning agent near the closed frontier on some benchmarks | Moonshot |
| 2025-11-13 | GPT-5.1 and GPT-5.1-Codex (API) | Faster, more steerable GPT-5 | OpenAI changelog |
| 2025-11-18 | Gemini 3 Pro and Google Antigravity | Topped LMArena at 1501 Elo; Google's agentic IDE | Google, Antigravity |
| 2025-11 | AI2 OLMo 3 (fully open) | Best fully open model family: data, code and checkpoints | AI2 |
| 2025-11-24 | Claude Opus 4.5 | Opus price cut to $5 / $25 | Anthropic |
| 2025-12-09 | Agentic AI Foundation formed (Linux Foundation): MCP, goose, AGENTS.md | Agent standards move to neutral governance | Linux Foundation, Anthropic |
| 2025-12-11 | GPT-5.2 | OpenAI's year-end flagship | OpenAI |
2026: the year of monthly releases¶
| Date | Event | Why it matters | Source |
|---|---|---|---|
| 2026-01-01 | Steve Yegge, "Welcome to Gas Town" (eight stages of AI-assisted development) | A widely shared map of how developers move from autocomplete to agent fleets | Medium |
| 2026-01 | Claude Cowork (research preview) | Claude Code-style agent for non-coding knowledge work | Claude |
| 2026-01-14 | Cursor reports hundreds of agents building a browser from scratch | Showcase and stress test of long-running multi-agent coding; critics found quality gaps | Cursor |
| 2026-01 | Kimi K2.5 (open) | Moonshot's multimodal agentic update | Hugging Face |
| 2026-01 (late) | Clawdbot, renamed Moltbot, becomes OpenClaw (reported) | Open-source personal agent that runs on your machine and talks through chat apps | GitHub |
| 2026-02-05 | Claude Opus 4.6 (1M context beta) | Start of Anthropic's rapid 2026 cadence | Anthropic |
| 2026-02-05 | GPT-5.3-Codex in Codex (API Feb 24) | OpenAI's coding-agent model | OpenAI changelog |
| 2026-02-05 | Anthropic: 16 parallel Claude agents build a C compiler | Concrete data on multi-agent cost (~$20K) and limits | Anthropic |
| 2026-02-11 | Z.ai GLM-5 (open, MIT) | Leading open coding model at release | Hugging Face |
| 2026-02-15 | OpenClaw's creator Peter Steinberger joins OpenAI; project moves to a foundation | Big labs compete for agent-tool builders | Steinberger |
| 2026-02-17 | Claude Sonnet 4.6 | Mid-tier refresh at $3 / $15 | Anthropic |
| 2026-02-19 | Gemini 3.1 Pro (preview) | Google's last Pro release before Gemini 4 | |
| 2026-02-23 | Anthropic names DeepSeek, Moonshot and MiniMax in a distillation report | ~24,000 fraudulent accounts and 16M+ exchanges; distillation becomes a policy issue | Anthropic |
| 2026-02-24 | METR productivity update | Later data suggests speedups, with large selection-effect caveats | METR |
| 2026-03-05 | GPT-5.4 and GPT-5.4 Pro | OpenAI flagship refresh | OpenAI changelog |
| 2026-03-09 | Anthropic's managed Code Review (research preview) | Multi-agent PR review as a product | Claude |
| 2026-04-02 | Google Gemma 4 (Apache 2.0) | Gemma switches from custom terms to Apache 2.0 | AI Studio changelog, Hugging Face |
| 2026-04-07 | Anthropic's Project Glasswing and Claude Mythos Preview (restricted) | A model held back for cyber defenders only; start of gated tiers | Anthropic |
| 2026-04-08 | Meta Muse Spark, the first Meta Superintelligence Labs model (closed) | Meta moves away from open Llama for its best models | Simon Willison |
| 2026-04-16 | Claude Opus 4.7 | Better vision and long-running work; new tokenizer; xhigh effort |
Anthropic |
| 2026-04-17 | Claude Design (Anthropic Labs research preview) | A closed product that combines model, harness and tools for visual work | Anthropic |
| 2026-04-23 | GPT-5.5 (API Apr 24) | OpenAI flagship refresh | OpenAI changelog |
| 2026-04-24 | DeepSeek V4 preview: V4-Pro (1.6T) and V4-Flash, 1M context | Open models reach 1M context by default | DeepSeek |
| 2026-05-19 | Google I/O: Gemini 3.5 Flash; 3.5 Pro promised for June | 3.5 Pro was later cancelled | |
| 2026-05-28 | Claude Opus 4.8 | Last Opus 4.x; fallback model for Fable's safety routing | Anthropic |
| 2026-06 | MiniMax M3 (open; non-commercial by default) | Example of a license tightening between versions | Hugging Face |
| 2026-06 | NVIDIA Nemotron 3 Ultra (550B open MoE) | Largest US open model at the time | Hugging Face |
| 2026-06-09 | Claude Fable 5 and Mythos 5: a new tier above Opus | First generally available Mythos-class model; $10 / $50 | Anthropic |
| 2026-06-12 | Anthropic suspends access to Fable 5 and Mythos 5 for all users after US export controls take effect | Frontier access can disappear for regulatory reasons; controls lifted Jun 30, Fable 5 back Jul 1 | Anthropic |
| 2026-06-30 | Claude Sonnet 5 at $2 / $10 | Sonnet price cut; near-Opus coding | Anthropic |
| 2026-07-09 | GPT-5.6 Sol, Terra and Luna | OpenAI adopts named tiers | OpenAI changelog |
| 2026-07-09 | Meta Muse Spark 1.1 and the Meta Model API (public preview) | First time developers can build on a Muse model | Meta |
| 2026-07-16 | Kimi K3 API; 1.56 TB of weights followed Jul 27 | 2.8T-parameter open model with a revenue-tiered license; frontier-scale weights no individual can run | Hugging Face, Simon Willison |
| 2026-07-16 | Hugging Face discloses an intrusion; OpenAI confirms Jul 21 its test agent did it | An autonomous agent escaped its sandbox during a cyber benchmark | Simon Willison |
| 2026-07 | Thinking Machines Lab releases Inkling (open, Apache 2.0) | New US lab ships a ~1T-parameter open MoE | Hugging Face |
| 2026-07-21 | Gemini 3.6 Flash and 3.5 Flash-Lite | Google keeps shipping Flash while Pro stalls | AI Studio changelog |
| 2026-07-24 | Claude Opus 5 | Led benchmarks at launch; communication style became the top complaint | Anthropic |
| 2026-08-03 | Alibaba Qwen3.8-Max (2.4T; weights about a week later) | First Max-tier Qwen with downloadable weights | The Batch, Hugging Face |
| 2026-08-05 | Demis Hassabis steps back as Google DeepMind CEO | Leadership shake-up amid Gemini delays | Fortune |
| 2026-08-10 | Meta Muse Glimmer 30B (open, Apache 2.0) | Meta's first open Muse model, built to run locally | Hugging Face |
| 2026-08-18 | Z.ai GLM-5.3 (weights Aug 25) | Top open-weight model on Epoch's cyber index | Hugging Face |
| 2026-09-01 | Claude Fable 5.1 and Mythos 5.1 | Top tier refresh; cache reads cut 75%; new anti-distillation rules | Anthropic, SD Times |
| 2026-09-02 | Gemini 3.8 Flash | Google's fourth Flash release since May (3.7 Flash came out August 13) | AI Studio changelog |
| 2026-09-03 | GPT-6 Astra | OpenAI's new top tier at $10 / $50 | OpenAI changelog |
| 2026-09-10 | DeepSeek V4.1-Flash | New architecture; replaces V4-Flash | DeepSeek |
| 2026-09 | xAI Grok 4.7 | xAI flagship at $2 / $6 | xAI |
| 2026-09-22 | Claude Opus 5.5 ($4 / $20) and, about 90 minutes later, GPT-6 Sol and Luna | A same-day "price war"; Opus 5.5 addressed Opus 5's communication complaints | Anthropic, TechCrunch |
| 2026-09-28 | Claude Sonnet 5.5 | Matches Opus 5.5 on several benchmarks at half the price | Anthropic |
| 2026-09-29 | GPT-6.1 Sol | Workhorse update a week after GPT-6 Sol | OpenAI changelog |
| 2026-09-30 | Google announces Gemini 4 Argon for cyber defenders first; drops Gemini 3.5 Pro | The delayed Gemini 4 flagship arrives through a staged rollout | Reuters via Khaleej Times |
| 2026-10-06 | Mistral Large 4 preview (~1T open MoE; weights Oct 29) | Strongest Western open model at launch, "about 6 months behind the frontier" | Mistral, Simon Willison |
| 2026-10-07 | Claude Haiku 5.5 at $0.10 / $0.50 | Small-model prices fall about 90% | Anthropic |
| 2026-10-07 | GPT-6 reaches ChatGPT with "Intelligent UI" | Answers can include interactive charts, forms and small tools | 9to5Mac |
| 2026-10-12 | Meta Muse Wonder listed (closed, $4 / $20) | Meta's frontier reasoning model reaches developers | OpenRouter |
| 2026-10-13 | Gemini 4 Argon generally available | Google's flagship back near the frontier, by its own benchmarks | |
| 2026-10-15 | Meta Muse Spark Open (816B open MoE, Muse Model License) | Meta's largest open release since Llama 4 | Hugging Face |
| 2026-10-15 | Anthropic's managed Code Review generally available | AI review moves from preview to product | Claude |
| 2026-10-16 | Epoch: open models trail closed by 6–7 months on cyber | Updated estimate of the open-closed gap | Epoch AI |
| 2026-10 | Reflection AI Beam 1.0 (501B open MoE, Apache 2.0) | Another US open-weight frontier attempt | Hugging Face |
| 2026-10-22 | Anthropic donates Agent Skills to the Agentic AI Foundation | Skills join MCP and AGENTS.md under neutral governance | Anthropic |
| 2026-10-23 | Kimi K3.1 Pro API; weights posted to Hugging Face in early November | The open-weight frontier as of this writing, priced at half of Opus 5.5 | Hugging Face, Let's Data Science |
| 2026-10-26 | GPT-6.1 Terra ($1 / $5) | Cheaper mid-tier aimed at computer use | OpenAI changelog |
| 2026-10-26 | Anthropic files an S-1 for a proposed IPO | A frontier lab heads to public markets | Anthropic |
| 2026-10-28 | Google announces Gemini 4 Carbon, first to vetted cyber defenders | Google adopts the restricted-frontier pattern Anthropic used for Mythos |
Patterns worth noticing¶
- The release interval shrank. Big launches came about yearly in 2020–2023 and quarterly in 2024–2025. By 2026 each top lab shipped something roughly every month.
- Open–closed gaps depend on the task. Epoch estimated about a year across its 2024 benchmark sample and 6–7 months on cyber tasks in 2026; these are different measures.
- Leadership in open weights moved to China after DeepSeek-R1 in January 2025. US and European labs answered in 2025–2026 with gpt-oss, Gemma 4, Nemotron, Inkling, Muse and Mistral Large 4.
- Agent tooling standardized. MCP (2024), AGENTS.md (2025) and Agent Skills (2025) all ended up at the Agentic AI Foundation.
- Incidents are part of the history. Sycophancy rollbacks, infrastructure bugs, suspended launches and an agent that escaped its sandbox all happened at the frontier labs themselves. See module 11 and module 14.