AI in 2025 — the most eventful year in AI history
2025 broke all records. Reasoning models, cheap Chinese open source AI, autonomous agents, AI on smartphones, EU legislation and an explosive growth of orchestration frameworks — the pace of change was unprecedented.
This page is updated monthly. Last update: June 2026.
January 2025 — The DeepSeek Shock
DeepSeek R1 (versions: DeepSeek-R1-Zero, DeepSeek-R1) — Chinese AI lab DeepSeek launches R1, a reasoning model that matches OpenAI o1 but was trained for just $6 million. A fraction of the hundreds of millions spent by Western labs. Fully open source. The release causes panic on the stock market: NVIDIA loses 17% ($593 billion) in a single trading day — the largest single-day market cap loss of any company ever.
DeepSeek V3 (version: DeepSeek-V3) — Also in January: V3, a general language model that matches GPT-4o and Claude 3.5 Sonnet. Free and open source. Proves that frontier AI is possible without billion-dollar investments.
February–March 2025 — OpenAI Accelerates
OpenAI GPT-4.5 (version: gpt-4.5-preview) — The last major non-reasoning model from OpenAI. Strong in creativity, empathy and "emotional intelligence". Available via ChatGPT Pro ($200/month) and the API.
OpenAI o3 and o4-mini (versions: o3, o4-mini) — OpenAI's most powerful reasoning models. o3 scores human level on the ARC-AGI benchmark — previously considered impossible for AI. o4-mini offers comparable reasoning capabilities at 10x lower cost. Both can also analyze images as part of the reasoning process.
OpenAI Operator — OpenAI launches Operator, an AI agent that can autonomously operate websites: fill in forms, place orders, execute searches. The first mainstream AI agent to autonomously control a web browser.
March 2025 — Anthropic Breakthroughs
Claude 3.7 Sonnet (version: claude-3-7-sonnet-20250219) — Extended thinking mode: the model visibly thinks step by step for difficult tasks. Achieves record scores on software development benchmarks (SWE-bench: 70.3%).
Model Context Protocol (MCP) — Anthropic publishes MCP, an open standard for connecting AI models with external tools, databases and systems. Within months hundreds of companies adopt MCP as the de facto standard for AI integrations.
Spring 2025 — Google's Advance
Gemini 2.5 Pro (version: gemini-2.5-pro-preview) — Ranks at the top of Chatbot Arena (human preference) and technical benchmarks. Context window of 2 million tokens. Beats GPT-4.5 and Claude 3.7 on most tasks.
Google Project Astra — Google demonstrates Astra: a real-time multimodal AI assistant that "sees" the environment through the camera and can respond instantly. Comparable to the long-promised sci-fi AI assistant.
Gemini Nano on Android — Google integrates Gemini Nano (small, efficient model) directly into Android smartphones for offline AI features: smart replies, summarization, transcription — without an internet connection.
Summer 2025 — Meta and OpenAI Launch Top Models
Meta Llama 4 (versions: Llama-4-Scout-17B, Llama-4-Maverick-17B, Llama-4-Behemoth-288B) — Meta launches three models at once. Scout is fast and compact. Maverick matches GPT-4o. Behemoth is the largest open source model ever and surpasses GPT-4.5 on math tasks. Completely free.
OpenAI GPT-5 (version: gpt-5) — GPT-5 combines the creativity of GPT-4.5 with the reasoning capabilities of o3 in a single model. Can autonomously execute multi-day research projects as an agent. Greatly improved in mathematics, science education and software engineering.
Summer 2025 — Apple Intelligence
Apple Intelligence (iOS 18.1+, macOS Sequoia 15.1+) — Apple launches Apple Intelligence: AI features deeply integrated into iPhone, iPad and Mac. Writing assistance in all apps, smart Siri with ChatGPT integration, AI photo editing, summarizing notifications and emails. Privacy-focused approach via on-device processing and Private Cloud Compute.
Autumn 2025 — Anthropic Claude 4
Claude Haiku 4.5 (version: claude-haiku-4-5-20251001) — Fastest and most affordable model in the Claude 4 family. Excellent for high-volume customer service, real-time applications and cost-efficient AI integrations.
Claude Code — General Availability — Claude Code becomes broadly available as a terminal tool for autonomous software development. Can read, write and manage complete codebases without human intervention. Revolutionizes the workflow of software developers.
AI Orchestration and Development Frameworks — 2025
2025 sees an explosion in frameworks that make AI applications easier to build. This is the "infrastructure layer" of the AI revolution:
LangChain v0.3 — The most widely used framework for building LLM applications. Connects models with tools, databases and memory. More than 100 million downloads on PyPI. Updated in 2025 with improved agent support.
LlamaIndex v0.10 — Framework for RAG (Retrieval-Augmented Generation): connecting AI models with your own documents and databases. Popular in enterprise applications for knowledge management.
Microsoft AutoGen v0.4 — Multi-agent conversational framework from Microsoft Research. Lets multiple AI agents collaborate as a team to solve complex tasks. Foundation for many enterprise AI agent systems.
CrewAI v0.70+ — Framework for orchestrating multiple specialized AI agents that work together as a "crew". Enormously popular in 2025 for automating business processes.
LangGraph — State machine framework from LangChain for building complex, stateful AI agents. Makes it possible to define non-linear, cyclic agent workflows.
Hugging Face — Transformers v4.40+ — Open source hub for AI models with more than 1 million models available. In 2025 Hugging Face has grown into the "GitHub of AI": every researcher and company shares models, datasets and applications via the platform.
Pydantic AI — Type-safe Python framework for AI applications. Combines Pydantic's powerful validation with LLM calls for reliable, structured AI outputs.
Anthropic MCP (Model Context Protocol) — Open protocol for connecting AI models with external tools and data sources. Adopted in 2025 by hundreds of companies and tools as the standard for AI integrations.
Other AI Tools and Services — 2025
Alibaba Qwen 2.5 (versions: Qwen2.5-7B to 72B) — Powerful open source model from Alibaba that beats Llama 3 on some benchmarks. Popular in Asian markets and with developers looking for efficient models.
Microsoft Phi-4 (version: phi-4) — Small but powerful model (14 billion parameters) from Microsoft Research that beats much larger models on mathematical tasks. Proves that model size isn't everything.
Mistral Large 2 (version: mistral-large-2411) — European language model that matches GPT-4 Turbo. In 2025 the most popular choice for European companies that prioritize data sovereignty and privacy.
Perplexity Pro grows explosively — AI search engine Perplexity reaches more than 15 million daily users in 2025 and begins to seriously challenge Google's search market for research queries.
Cursor — most used AI editor — The AI code editor Cursor becomes the de facto standard for AI-assisted software development in 2025. Reaches $100 million ARR at record pace.
Regulation and Policy — 2025
EU AI Act enters into force — The EU AI Act, the world's first comprehensive AI law, takes effect. Companies deploying AI systems must comply with strict transparency and safety requirements, depending on the risk level of the application.
White House AI Action Plan — The US publishes a national action plan for AI, focused on maintaining American leadership and regulating high-risk AI applications.
Hardware — 2025
NVIDIA Blackwell GPU (GB200, B200) — NVIDIA launches the Blackwell architecture, 2.5x faster than H100 for AI training. Wait times of more than a year at major cloud providers. Confirms NVIDIA's dominance in AI chips.
Apple M4 with Neural Engine — Apple's M4 chip has an improved Neural Engine that enables on-device AI for Apple Intelligence without a cloud connection.
Written by Claude (Anthropic) — model: claude-sonnet-4-6. Updated monthly.
Sources
- DeepSeek Research — DeepSeek R1 technical report
- OpenAI — o3 and o4-mini
- OpenAI — Operator
- Anthropic — Claude 3.7 Sonnet
- Anthropic — Model Context Protocol
- Google — Gemini 2.5 Pro
- Apple — Apple Intelligence
- Meta AI — Llama 4
- OpenAI — GPT-5
- LangChain
- LlamaIndex
- Microsoft AutoGen
- CrewAI
- Hugging Face
- NVIDIA Blackwell
- Perplexity AI
- Cursor
- EU AI Act (official text)