Breakthroughs and who contributed to each wave
OpenAI releases GPT-5.4 with 1M API context, native computer-use for agents, and 33% fewer factual errors. Available in standard, Thinking, and Pro variants.
Anthropic releases auto mode for Claude Code — executes file writes and bash without per-action approval. AI safety classifier reviews every tool call. Claude Code drives ~4% of all public GitHub commits.
OpenAI shuts down Sora/Sora 2 just six months after its #1 App Store debut, citing unsustainable costs. The planned $1B Disney partnership collapses.
Anthropic releases Opus 4.6 with 1M token context window, Agent Teams (multi-agent collaboration), effort controls for speed/cost tradeoffs, and 14.5-hour agentic task completion.
Global VC hits $189B in February — the largest startup funding month ever. OpenAI raises $110B at $840B valuation. Anthropic raises $30B at $380B. Three companies absorb 83% of all global VC.
Google releases Gemini 3.1 Pro succeeding both 2.5 Pro and 3 Pro. Becomes Google's recommended frontier model.
Cursor announces agents that test their own changes, record work through video/screenshots, and run in parallel on VMs. Adds automations triggered by Slack, GitHub, and PagerDuty.
xAI raises $20B ($7.5B equity) in Series E. Expands Colossus to 555,000 GPUs across 2GW. One of the largest private rounds ever.
California Transparency in Frontier AI Act and Texas Responsible AI Governance Act take effect. South Korea Basic AI Act enters force. First comprehensive AI-specific laws globally.
Mistral releases 10 models including Large 3 (41B active / 675B MoE) for frontier reasoning and Ministral 3 (3B/7B/14B) for edge. Competitive with much larger labs.
OpenAI releases GPT-5.2, capping an extraordinary 25-day period of four frontier model releases across the industry.
MCP donated to the Agentic AI Foundation (AAIF) under Linux Foundation. Co-founded by Anthropic, Block, and OpenAI with support from Google, Microsoft, AWS, and Cloudflare.
Google releases Gemini 3, achieving breakthrough scores on Humanity's Last Exam, GPQA Diamond, and MathArena Apex (23.4%). Tops the LMArena Leaderboard.
MCP receives major updates: async operations, statelessness, server identity, official registry. Over 5,800 MCP servers and 300+ clients. 97M+ monthly SDK downloads.
Anthropic releases Opus 4.5 with major improvements in coding and workplace tasks. Part of the historic four-frontier-model November blitz.
Black Forest Labs releases FLUX 2 Dev, a 32B-parameter open-source image model using Rectified Flow Transformers. Built by ex-Stable Diffusion team.
Physical Intelligence raises $600M led by Alphabet's CapitalG with Bezos participating. Develops AI for generalizable robot learning.
OpenAI officially restructures as a public benefit corporation valued at ~$500B. Nonprofit OpenAI Foundation holds ~26% ($130B), Microsoft ~27% ($135B). Paves way for IPO.
Anthropic closes a $13B Series F led by ICONIQ, reaching $183B post-money valuation. Reportedly captures ~40% of enterprise LLM spend.
Mistral AI raises €1.7B in Series C with ASML contributing €1.3B. NVIDIA also participates. Valuation reaches €11.7B.
xAI's Colossus supercomputer in Memphis expands to 200,000+ GPUs with plans for 555,000. The largest single-site AI training installation in the world.
OpenAI releases GPT-5 with a smart router switching between fast and deep reasoning modes. 94.6% on AIME 2025, 74.9% on SWE-bench, ~80% fewer factual errors. Sets new industry benchmarks across the board.
OpenAI launches two open reasoning models — a shocking strategic shift toward open source under competitive pressure from DeepSeek and Meta.
Anthropic releases Opus 4.1 with improved code generation, search reasoning, and instruction adherence.
Apple announces at WWDC that its on-device foundation models (~3B params, 2-bit quantized) will be available to third-party apps via Foundation Models API with tool calling support.
Anthropic releases Claude Opus 4 and Sonnet 4 with extended thinking. Claude Code goes GA. Opus 4 described as "the world's best coding model" — works autonomously for hours.
OpenAI launches Codex — autonomous coding agent powered by codex-1 (optimized o3). Each task gets a cloud sandbox preloaded with the user's repo. Reaches 1M+ weekly users.
Google releases Veo 3 at I/O 2025 with synchronized dialogue and sound effects for video. Project Astra updates add memory and computer control. Project Mariner brings computer-use to Gemini API.
DeepSeek releases R1-0528 with enhanced reasoning through more compute and advanced post-training. Proves DeepSeek is not a one-hit wonder.
OpenAI releases GPT-4.1, 4.1 mini, and 4.1 nano — developer models with 1M token context and 21% better coding. GPT-4.5 deprecated. Aggressive pricing at $2/M input tokens.
Full o3 reasoning model and o4-mini launch. O3 is the first reasoning model to agentically use all ChatGPT tools (web search, Python, image gen, file analysis). 20% fewer errors than o1.
Meta releases Llama 4 with Scout and Maverick variants using Mixture-of-Experts. Natively multimodal with vision, video, and text in a single forward pass. Largest open-source model release.
Cognition releases Devin 2.0 with Interactive Planning and Devin Wiki. Price drops from $500/month to $20/month + $2.25/ACU, making autonomous AI coding accessible to individuals.
Midjourney V7 launches after nearly a year with dramatically improved photorealism, emotional depth, cinematic lighting, and a 10x faster Draft Mode.
Google releases its first purpose-built "thinking model" with 2M token context, native multimodality, and Deep Think mode. Leads benchmarks in math (AIME 2025) and science (GPQA Diamond).
OpenAI closes $40B funding led by SoftBank ($30B) — the largest private tech funding round ever at the time. Microsoft, Coatue, and Thrive participate.
Jensen Huang unveils Blackwell Ultra (1.5x faster), Vera Rubin family (3x Blackwell, 2026), Rubin Ultra (14x, 2027), and Feynman architecture (2028). $1T in demand projected through 2027.
OpenAI officially adopts Anthropic's Model Context Protocol across products including the ChatGPT desktop app. MCP becomes the de facto industry standard when its biggest competitor adopts it.
First enforcement phase of the EU AI Act bans social scoring, emotion inference in workplaces, real-time biometric ID, and manipulative AI. Penalties up to €35M or 7% global turnover.
xAI releases Grok 3 trained with 10x more compute than Grok 2 on the Colossus supercomputer (200K GPUs). Musk calls it "the smartest AI on Earth."
Anthropic releases Claude Code as a terminal-native agentic coding tool. Later reaches $1B+ annualized revenue within 6 months and drives ~4% of all public GitHub commits.
OpenAI releases GPT-4.5 ("Orion") — its largest pre-trained model, emphasizing EQ and reduced hallucination. Called OpenAI's "last non-chain-of-thought model." Later deprecated for GPT-4.1.
Elon Musk leads a consortium offering $97.4B to buy OpenAI's nonprofit arm. The board unanimously rejects: "OpenAI is not for sale." A federal judge denies Musk's attempt to block OpenAI's for-profit conversion.
Chinese startup DeepSeek releases R1 (671B params, 37B active via MoE) matching OpenAI o1 on reasoning benchmarks. Trained for under $6M. Released under MIT license. Overtakes ChatGPT as #1 free app on App Store.
DeepSeek R1's cost efficiency revelation triggers the largest single-day market cap loss in US stock history. NVIDIA drops ~$600B. Nasdaq plunges 3%+. Challenges the assumption that AI progress requires massive capital.
President Trump announces the Stargate Project — a $500B AI infrastructure venture led by SoftBank and OpenAI. $100B to be deployed immediately. The largest AI infrastructure investment ever.
OpenAI releases o3-mini to all ChatGPT users including the free tier, democratizing access to chain-of-thought reasoning for the first time.
OpenAI launches the full o1 reasoning model and ChatGPT Pro ($200/month tier) on Day 1 of "12 Days of Shipmas." Sora video generation launches publicly on Day 3.
Google releases Gemini 2.0 Flash with Multimodal Live API, native image generation, and integrated search. Outperforms Gemini 1.5 Pro at 2x the speed.
OpenAI announces the o3 reasoning model family on the final day of Shipmas, claiming near-AGI performance on ARC-AGI benchmark.
Anthropic introduces MCP as an open standard for connecting AI assistants to external tools and data. Released with SDKs and pre-built servers for GitHub, Slack, Google Drive, and Postgres.
Geoffrey Hinton and John Hopfield awarded the Nobel Prize in Physics for foundational discoveries enabling machine learning with artificial neural networks — the first Nobel for AI research.
Demis Hassabis and John Jumper win the Nobel Prize in Chemistry for AlphaFold2, which solved the 50-year protein structure prediction problem. David Baker shares the other half for computational protein design.
Anthropic releases upgraded Claude 3.5 Sonnet with "computer use" — the AI can control a desktop by interpreting screenshots and simulating mouse/keyboard input. First major AI agent controlling real GUIs.
