OpenAI's latest frontier model—SOTA on most major benchmarks, with standout strength in agentic coding, computer use, and knowledge work.

OpenAIopens in a new tab

A SOTA model for advanced autonomous cybersecurity reasoning, released restrictively for safety, alongside Project Glasswing to secure open-source software with major industry partners.

Anthropicopens in a new tab

A frontier model family for professional work and long-running agents, including Pro, Thinking, and base variants.

OpenAIopens in a new tab

Anthropic's most capable model brought a reported 20% accuracy gain, Infinite Chats, and stronger Excel and financial-modeling workflows.

Anthropicopens in a new tab

Google's most powerful model replaced the 2.5 series and claimed a 50%+ improvement over Gemini 2.5 Pro.

Google DeepMindopens in a new tab

A next-generation flagship with a major intelligence leap across coding, math, writing, and reasoning.

OpenAIopens in a new tab

OpenAI's first open-weight models since GPT-2 arrived in 20B and 120B variants, marking a historic open-source shift.

OpenAIopens in a new tab

xAI's most powerful model brought a major reasoning leap and was trained on an expanded Colossus cluster.

xAIopens in a new tab

A massive one-trillion-parameter MoE with open weights, competitive with frontier models and a major Chinese AI milestone.

Moonshot AIopens in a new tab

Anthropic's strongest model generation pushed parallel tool use and long autonomous tasks.

Anthropicopens in a new tab

An agentic coding tool that can write, test, and fix code end-to-end, integrated into ChatGPT for Pro, Team, and Enterprise, laying the foundation for today's Codex.

OpenAIopens in a new tab

Scout and Maverick brought native multimodality, with Scout reaching a 10-million-token context window.

Meta AIopens in a new tab

Gemini 2.5 Pro launched with built-in reasoning, a one-million-token context window, and the top LMArena position at launch.

Googleopens in a new tab

An agentic coding tool in the terminal that can read, modify, test, and execute across entire codebases, laying the foundation for today's Claude Code.

Anthropicopens in a new tab

An open reasoning model rivaling o1, built around a pure RL approach and strong enough to send shockwaves through the market.

DeepSeekopens in a new tab

A 671B-parameter MoE trained for a reported $5.5M and competitive with GPT-4o and Claude 3.5 Sonnet, reshaping the cost-efficiency debate.

DeepSeekopens in a new tab

Sora 1 set a new standard for text-to-video with strong realism and motion, opening broad access and accelerating the race between OpenAI and rivals such as Google’s later Veo in generative video.

OpenAIopens in a new tab

An open-source protocol for connecting LLMs to tools, data, and systems, standardizing how models access external context across the AI ecosystem.

Anthropicopens in a new tab

The first OpenAI reasoning model put inference-time reasoning in the spotlight and targeted PhD-level science and math tasks.

OpenAIopens in a new tab

Competitive with GPT-4o and Claude 3.5 and available on the X platform.

xAIopens in a new tab

The 405B model became the largest open model of its moment, with 128K context and performance near GPT-4 on many benchmarks.

Meta AIopens in a new tab

Claude 3.5 Sonnet surpassed GPT-4o and Gemini 1.5 Pro on several benchmarks while running twice as fast as Claude 3 Opus at lower cost.

Anthropicopens in a new tab

An omni model with native audio, vision, and text that was twice as fast and 50% cheaper than GPT-4 Turbo.

OpenAIopens in a new tab

Trained on 15 trillion tokens and released in 8B and 70B sizes, LLaMA 3 became a new open-source state of the art with massive adoption.

Meta AIopens in a new tab

xAI's first open-source model was a 314B-parameter MoE released under Apache 2.0 and among the largest open MoE models at the time.

xAIopens in a new tab

Haiku, Sonnet, and Opus brought 200K context and vision, with Opus matching GPT-4 on many benchmarks.

Anthropicopens in a new tab

A one-million-token context window, MoE architecture, and whole-codebase processing made it a major leap in long-context modeling.

Googleopens in a new tab

Google's Gemini family arrived in Nano, Pro, and Ultra sizes and was designed as natively multimodal from training.

Googleopens in a new tab

One of the first major open-weight model releases for commercial use, with 7B, 13B, and 70B sizes plus RLHF-tuned chat variants.

Meta AIopens in a new tab

Claude 2 expanded Anthropic's public model lineup with stronger capabilities and broader access through chat and API surfaces.

Anthropicopens in a new tab

Claude and Claude Instant became Anthropic's first broadly introduced assistant models and a direct ChatGPT competitor.

Anthropicopens in a new tab

GPT-4 could accept both text and images, raising the bar for frontier-model reasoning and multimodal capabilities.

OpenAIopens in a new tab

Elon Musk founded xAI with a mission to develop AI that advances our understanding of the universe.

The Vergeopens in a new tabxAIopens in a new tab

A very early version of Cursor put AI-native chat, edits, and codebase reasoning inside the IDE; it later grew into one of the leading developer tools and helped kickstart vibe coding.

Cursoropens in a new tab

LLaMA 1 and its leaked weights ignited the open-source LLM wave and showed smaller models could approach GPT-3-class performance.

Meta AIopens in a new tab

GPT-3.5 with RLHF in a chat interface reached 100 million users in two months and helped define the modern AI era.

OpenAIopens in a new tab

Proved that smaller models trained on more data can outperform larger undertrained ones, redefining scaling laws.

Google DeepMindopens in a new tab

Introduced RLHF for alignment and pioneered the technique of training models to follow human instructions safely.

OpenAIopens in a new tab

Dario and Daniela Amodei and a group of former OpenAI colleagues (reported as seven co-founders, including Jared Kaplan, Jack Clark, and Chris Olah) form Anthropic as a public benefit corporation, prioritizing safety after strategic tension with OpenAI’s direction on speed and commercialization.

Anthropicopens in a new tab

A 12-billion-parameter transformer trained on text–image pairs to synthesize images from natural-language captions—combining disparate concepts, controlling attributes and viewpoint, and inpainting regions—showing that rich visual compositions can be steered with words.

OpenAIopens in a new tab

OpenAI releases GPT-2 but withholds the largest version at launch due to misuse concerns. The decoder-only transformer is trained with next-token prediction to generate coherent long-form text.

OpenAIopens in a new tab

OpenAI reveals the first Generative Pre-trained Transformer (GPT), demonstrating unsupervised pre-training followed by supervised fine-tuning for strong downstream NLP performance.

OpenAIopens in a new tab

Facebook releases PyTorch, a Python-first deep learning framework that would eventually become the dominant framework for AI research.

PyTorchopens in a new tab

Announced as a nonprofit with co-chairs Sam Altman and Elon Musk, CTO Greg Brockman, research director Ilya Sutskever, and founding researchers including Andrej Karpathy, John Schulman, Durk Kingma, and Wojciech Zaremba—committed to digital intelligence that benefits humanity and broad publication of results.

OpenAIopens in a new tab

Google open-sources TensorFlow, its internal deep learning framework. Initially developed by the Google Brain team, TensorFlow would become one of the most influential AI frameworks.

Google Researchopens in a new tab

Demis Hassabis, Shane Legg, and Mustafa Suleyman start DeepMind to pursue general AI via deep reinforcement learning and games; Google acquired the lab in 2014, and in April 2023 it merged with Google Brain as Google DeepMind.

Google DeepMindopens in a new tab