2025 was a landmark year in the evolution of large language models, marked by rapid advancements across reasoning, agent systems, tool integration, and open-source competition. The year began with OpenAI’s groundbreaking o1 and o1-mini models in September 2024, which introduced Reinforcement Learning from Verifiable Rewards (RLVR) as a dominant training paradigm. This approach, which encourages models to break down complex problems into intermediate steps and verify their own reasoning, became the defining trend of 2025. By early 2025, OpenAI followed up with o3, o3-mini, and o4-mini, while nearly every major AI lab released their own reasoning-optimized models. The result was a dramatic leap in problem-solving capability, particularly in math, code debugging, and multi-step planning. The real power of reasoning emerged not in isolated puzzles, but when combined with tool use. Reasoning models could now plan, execute, and refine actions across multiple steps—transforming AI from a passive responder into an active collaborator. This enabled practical breakthroughs in AI-assisted search, where models could now produce detailed, accurate reports in minutes, and in code development, where agents could identify and fix complex bugs through iterative execution and analysis. This shift gave rise to the year of agents, particularly in coding and research. The quiet February release of Claude Code by Anthropic marked a turning point. Though not announced with fanfare, it introduced a powerful asynchronous coding agent capable of writing, running, and refining code autonomously. OpenAI’s Codex Cloud, Google’s Jules, and other CLI-based agents like GitHub Copilot CLI and OpenHands CLI followed suit. These tools allowed developers to offload entire workflows—prompting, coding, testing, and PR generation—often from their phones, revolutionizing how software is built. The command line became a central interface for AI, with tools like Claude Code and Codex CLI proving that developers embrace LLMs in terminal environments when they’re powerful and reliable. The success of these tools, especially Claude Code, which reportedly generated $1 billion in annual recurring revenue by December, demonstrated that the CLI is not a niche but a mainstream workflow. A major cultural shift occurred with the normalization of risk. As agents ran in safe, isolated environments like cloud sandboxes, users began adopting “YOLO mode”—bypassing safety checks for faster results. Security researcher Johann Rehberger’s concept of the “normalization of deviance” warned that repeated safe outcomes could erode caution, creating long-term risks. This tension between speed and safety became a defining theme. Pricing also evolved. The $20/month ChatGPT Plus model remained popular, but a new tier emerged: $200/month subscriptions like Claude Pro Max 20x and OpenAI’s ChatGPT Pro. These plans became viable for heavy users, especially those running coding agents that consume vast token counts. The economic model shifted from pay-per-token to subscription-based access for high-volume workflows. China’s open-weight model revolution dominated the second half of the year. Models like GLM-4.7, Kimi K2 Thinking, DeepSeek V3.2, and MiniMax-M2.1 outperformed many Western models, with several ranking in the top five on Artificial Analysis’s open-weight leaderboard. The release of DeepSeek 3 and DeepSeek R1 in late 2024 sent shockwaves through the market, briefly causing a $600 billion drop in NVIDIA’s market cap as investors questioned the U.S. monopoly on AI. These models, often fully open-source under permissive licenses, pushed the state of the art in efficient training and inference. The year also saw LLMs achieve gold medals in elite academic competitions. In July, both OpenAI and Google Gemini scored gold in the International Math Olympiad without access to tools, proving that advanced reasoning can solve novel, complex problems. In September, they achieved similar success in the International Collegiate Programming Contest, using code execution environments. Meanwhile, OpenAI’s lead eroded. While still dominant in consumer mindshare, they were overtaken in image generation by Google’s Nano Banana models, in code by models like Claude Opus 4.5, and in open-weight benchmarks by Chinese labs. Google’s Gemini 3, powered by in-house TPUs, proved highly competitive in both performance and cost efficiency. The year also brought cultural phenomena. The “pelican riding a bicycle” challenge, initially a joke, became a viral benchmark for model capability. Its absurdity made it a fun, unofficial test of creativity and visual reasoning. Similarly, the term “slop” was named Merriam-Webster’s Word of the Year, capturing public concern over low-quality AI-generated content. Despite progress, risks remain. The integration of LLMs into browsers—like OpenAI’s ChatGPT Atlas and Google’s Gemini in Chrome—raises serious security concerns, particularly around prompt injection and data exfiltration. The “lethal trifecta” concept helped clarify the most dangerous form of such attacks: when an agent is tricked into stealing private data. Finally, the year highlighted the power of conformance suites—testable, language-agnostic standards that allow models to learn and verify new protocols. This approach may help new technologies gain traction without relying solely on training data. In summary, 2025 was the year LLMs stopped being toys and became tools of real, productive work. From reasoning and agents to open-source breakthroughs and new workflows, the landscape evolved faster than ever. The future is not just smarter models—it’s smarter ways of using them.